Dynamical Cognitive Science (Bradford Books)

Dynamical Cognitive Science

This page intentionally left blank

Dynamical Cognitive Science

A Bradford Book The MIT Press Cambridge, Massachusetts London, England

Lawrence M. Ward

( 2002 Massachusetts Institute of Technology All rights reserved. No part of this book may be reproduced in any form by any electronic or mechanical means (including photocopying, recording, or information storage and retrieval) without permission in writing from the publisher. This book was set in Times New Roman in 3B2 by Asco Typesetters, Hong Kong, and was printed and bound in the United States of America. Library of Congress Cataloging-in-Publication Data Ward, Lawrence M. Dynamical cognitive science / Lawrence M. Ward. p. cm. ``A Bradford book''. Includes bibliographical references and index. ISBN 0-262-23217-0 (hc. : alk. paper) 1. Cognitive science. 2. Change (Psychology). 3. TimeÐPsychological aspects. I. Title. BF311 .W2695 2001 153Ðdc21 2001044336

Contents

Preface

xiii

Chapter 1 Magic, Ritual, and Dynamics

1

1.1

Magic and Ritual

1.2

Dynamics

2.1

The Serial Universe

1

4

Chapter 2 Sequence

9

9

2.2 The Problem of Serial Order in Behavior 11 2.3 13

Markovian Analysis of Behavior

3.1

The Dance of Life

3.2

Music and Rhythm

3.3

Rhythms in the Brain

4.1

Space-Time

4.2

The Arrow of Time

4.3

Measuring Time

Chapter 3 Rhythms of Behavior

17

17 19 20

Chapter 4 Time

25

25 28 30

vi

Contents

Chapter 5 Cognitive Processes and Time

35

5.1 Temporal Unfolding of Cognitive Behavior 35 5.2 40

Timing of Cognitive Behavior

5.3

Memory

41

6.1

Systems

46

6.2

General Systems Theory

6.3

Dynamical Systems Theory

7.1

The Mandala of Science

7.2

Formal Theories

7.3 59

Principle of Complementarity

8.1 61

Theories, Models, and Data

8.2

Statical Models

8.3

Dynamical Models

Chapter 6 Systems and General Systems Theory 45

48 50

Chapter 7 Science and Theory

53

53

56

Chapter 8 Dynamical versus Statical Models

61

62 64

8.4 Why We Need Both Statical and Dynamical Models 68 Chapter 9 Dynamical and Structural Models 71

9.1

Structural Models

9.2

Graph Theory

71

72

9.3 Interplay between Dynamical and Structural Models 78

Contents

vii

Chapter 10 Deterministic versus Stochastic Dynamical Models 81

10.1 Deterministic Models 10.2 Stochastic Models

81 84

10.3 Do We Need Both?

87

Chapter 11 Linear Time Series Analysis

89

11.1 Time Series and Noise 11.2 ARIMA ( p; d; q)

89

90

11.3 ARIMA Model of Time Estimation 94 11.4 Mixed Regression-ARIMA Model of Psychophysical Judgment 96 Chapter 12 Probability Theory and Stochastic Models 99

12.1 Dynamical Cognitive Science and Mathematics 99 12.2 Stochastic Processes: A Random Walk to Ruin 100 12.3 Critical Points in Stochastic Models 103 12.4 Ergodicity and the Markov Property 105

Chapter 13 Stochastic Models in Physics

107

13.1 The Master Equation 13.2 Quantum Physics

110

13.3 Complementarity Redux Chapter 14 Noise

115

14.1 What Is Noise?

107

115

113

viii

Contents 14.2 Probabilistic Description of Noise 119 14.3 Spectral Properties of Noise 122

Chapter 15 Colored Noise

125

15.1 The Ubiquity of Colored Noise 125 15.2 The Vicissitudes of the Exponent a 129 15.3 Colored Noise in Living Systems 131

Chapter 16 1/f Noise in Human Cognition

135

16.1 Music and Time Perception 135 16.2 Reaction Time

138

Chapter 17 1/f Noise in the Brain

145

17.1

Neural Activity

145

17.2 Magnetoencephalogram Recordings 147 17.3 Electroencephalogram and Event-Related Potential Recordings 150 Chapter 18 Models of 1/f Noise

155

18.1 The Simplest Case 18.2 Multiplicative Noise

155 156

18.3 Self-Organized Criticality 18.4 Center-Surround Neural Network 160

157

Contents

ix

Chapter 19 Statistical Theory of 1/f Noise

165

19.1 What Must Be Explained 19.2 Queuing in a Wire 19.3 ARIMA (1; 0; 0)

165

166 169

19.4 Multifractals and Wild Self-A½nity 171 Chapter 20 Stochastic Resonance

175

20.1 What is Stochastic Resonance? 175 20.2 Stochastic Resonance in a Threshold Detector 179

Chapter 21 Stochastic Resonance and Perception 183

21.1 Detection of Weak Signals by Animals 183 21.2 Stochastic Resonance in Human Perception 188

Chapter 22 Stochastic Resonance in the Brain

193

22.1 Stochastic Resonance in Neurons 193 22.2 Neural Networks

196

Chapter 23 Chaos

201

23.1 Chaos Is Not What You Think It Is 201 23.2 What Chaos Really Is

203

23.3 Phase Space Drawings and Strange Attractors 207 Chapter 24 Chaos and Randomness

209

24.1 A Random Walk through the Logistic Di¨erence Equation 209

x

Contents 24.2 Dimensionality of an Attractor 213 24.3 Chaos and Noise

215

Chapter 25 Nonlinear Time Series Analysis

219

25.1 State Space Reconstruction 219 25.2 Out-of-Sample Forecasting 221 25.3 Surrogate Data

226

Chapter 26 Chaos in Human Behavior?

229

26.1 Could Unexplained Variance Be Chaos? 229 26.2 Nonlinear Forecasting Analysis of Time Estimation 230 26.3 Nonlinear Analysis of Mental Illness 233 26.4 Memory and the Logistic Di¨erence Equation 235

Chapter 27 Chaos in the Brain?

237

27.1 The Smell of Chaos

237

27.2 Dimensionality of the Electroencephalogram 240 27.3 Chaotic Event-Related Potentials? 242 Chapter 28 Perception of Sequence

245

28.1 The Gambler's Fallacy 28.2 Estimation of Short-Run Probabilities 247

245

Contents

xi 28.3 Evolution of Contingency Perception 251

Chapter 29 Can People Behave Randomly?

255

29.1 No!

255

29.2 Sometimes

258

29.3 Sequential Dependencies and Extrasensory Perception 261 Chapter 30 Can People Behave Chaotically?

263

30.1 Yes!

263

30.2 Not Really

265

30.3 Heuristics and Chaos

268

Chapter 31 Relaxation Oscillators: A Foundation for Dynamical Modeling 273

31.1 A Brief Taxonomy of Oscillators 273 31.2 The van der Pol Relaxation Oscillator 277 31.3 Noisy Oscillators in the Brain 280

Chapter 32 Evolution and Ecology of Cognition 285

32.1 Evolution of Cognition 32.2 Ecology of Cognition

Chapter 33 Dynamical Cognitive Neuroscience 295

33.1 Brain Imaging 33.2 Brain Dynamics 33.3 Hybrid Models

296 299 303

285 291

xii

Contents

Chapter 34 Dynamical Computation

305

34.1 Numerical Methods

305

34.2 Neural Network Models 34.3 Exotic Computers

312

Chapter 35 Dynamical Consciousness

315

35.1 Consciousness 35.2 Unity of Science

References Index

349

325

315 324

308

Preface

Life is change. How it di¨ers from the rocks. ÐJe¨erson Airplane, Crown of Creation The times they are a-changin'. ÐBob Dylan

Changing time, timely change, change creating time, time measuring changeÐthe themes of this book are change and time in various permutations and combinations. The book also deals with nonlinearity, chaos, randomness, and stochastic models, the use of computers to study complicated systems of di¨erential equations, systems theory, complementarity, the importance of formal models, methods from physics and mathematics for the analysis of cognitive systems, and interdisciplinarity, among other topics. Dynamical cognitive science is a potpourri of the old and the new, the borrowed and the ``true'' (in the carpenter's sense, meaning ``linear''). Many have proclaimed it was about time we began to focus on dynamics in cognitive scienceÐor in psychology as a whole. Gregson (1983) was perhaps the most forceful of the moderns, but you will ®nd many others referenced within, notably, van Gelder and Port (1995), Port and van Gelder (1995), Kelso (1995), van Gelder (1998), and Beer (2000). Their articles and books discuss the advantages and disadvantages of the dynamical approach to cognition and provide many examples of its usefulness in cognitive science. The present book is both more and less than these others. It is more because it discusses di¨erent topics in di¨erent ways, for example, noise and relaxation oscillators. It is less because it often touches on material covered in depth elsewhere and does not address at all approaches championed by other dynamical authors, for

xiv

Preface

example, the dynamical psychophysics of Gregson (1988, 1992) and the dynamical hypothesis of van Gelder (1998). Dynamical Cognitive Science is intended to be a deep introduction to the ®eld. Its purpose is to whet appetites, not to give a complete meal: although it treats usually neglected topics such as noise in depth, where it considers deep topics such as time and di¨erential equation models of cognition, it does so in interesting, rather than exhaustive ways. The book's thirty-®ve chapters are relatively short. Some introduce important topics presented elsewhere in detail, others together present an important topic in depth (e.g., chaps. 14±19 on noise). Each chapter is self-contained, making a few major points or introducing a few important topics amply yet succinctly, so that it can be read at a short sitting. Although I hope something in the book will appeal to a wide range of readers, from cognitive psychologists and neuroscientists to computer scientists, engineers, physicists, mathematicians, and philosophers, I do not expect that every reader will ®nd every chapter of interest. Thus mathematicians will probably want to skip over sections where familiar equations are explained, just as physicists will the introductory discussions of the master equation, quantum theory, and oscillators. On the other hand, psychologists will probably ®nd most of the chapters of interest, and much of the material new. Although some chapters are di½cult, there is really no serious mathematics in the book, only relatively simple equations that will become familiar as readers progress. I hope to show how the relevant mathematical forms (mostly di¨erence or di¨erential equations) can be approached in a more-or-less empirical way, especially by computer exploration, so that their meaning and usefulness to cognitive science is revealed. This book is much less than I had hoped it would be, or than it could have been had I spent ten more years working on it. The cascade of new dynamical work being undertaken by psychologists and other cognitive scientists promises to undercut the timeliness of one theme of this book, that dynamics should form a core approach to cognitive science. On the other hand, coming after decades of neglect, the new work makes a wideranging, comprehensible, yet deep introduction to the promise of the dynamical approach even more timely. I was helped by many people, not least by those whose work I have cited and adapted, and by countless other colleagues from whom I have learned over the past thirty years. I owe special thanks to Christian Richard, who read much of the text and provided astute comments; my graduate students Shuji Mori, Odie Geiger, Robert West, John

Preface

xv

McDonald, David Prime, and Matt Tata, who kept me thinking and put up with my obsessions; my other students and research assistants, especially Amanda Jane, Sharna Hurwitz, and Sarah Knowlden, who managed the references and the permissions, and Sim Desai and Dan Rootman, who helped get the stochastic resonance studies going; my colleagues and friends Cindy Greenwood, Birger Bergerson, Steve Marion, and other members of the Crisis Points Group at the University of British Columbia (UBC), from whom I learned an enormous amount, especially about stochastic dynamics; my colleague and friend Frank Moss, who ®rst inspired and then conspired in the study of stochastic resonance; my graduate school mentor and friend Gregory R. Lockhead; my wife, Brigitte Renahte Voegtle, who has endured the curse of the absent-minded professor for over ten years; and ®nally my mother, Sarah Mabel Johnston Ward Walls, who has always supported and encouraged me. Preparation of the book was funded by an ongoing operating grant from the Natural Sciences and Engineering Research Council (NSERC) of Canada, which has funded my research for over twenty-six years, and by a major thematic grant from the Peter Wall Institute for Advanced Studies at UBC to the Crisis Points Group, of which I am a member.

Chapter 1 Magic, Ritual, and Dynamics

In the magic of incantation, a formula of exorcism or of cursing possesses the virtue of subjugating to the will of him who pronounces it (in the right ``ritual'' conditions as to words, intonation, rhythm) beings and things in the world above and in the world here below. ÐMaurice Bouisson, Magic

1.1

Magic and Ritual

Why begin a book on dynamical cognitive science with a discussion of magic and ritual? Surely not to make an analogy between mystery and science. Surely not to suggest that magic and ritual are fertile ®elds to plow with mathematical models. And surely not to contend that our current understanding of these topics should form prototypes for the modeling process. Why, then? Because they dramatically illustrate the fundamental insight that, along with the rest of the universe, human behavior unfolds in time. This include both covert behaviors involved in sensation, perception, and cognition and the overt, observable motor behaviors from which we cognitive scientists must infer those covert behaviors. What, then, can we glean from magic and ritual to inspire us to study dynamical cognitive science? Magic and ritual are deeply embedded in the human condition. For much of history, magic and ritual ruled human behavior, and for many so-called primitive peoples, they still do. Magic, and ritual based on it, explained the mysteries of the world, informed religions, and helped humans cope with a powerful and uncaring natural world. Of the three forms of magic, sympathetic (like attracts like), incantational (see chapter epigraph), and talismanic (certain objects have magical power), incantational has been considered the most powerful: the words and numbers used in incantations are held to have real power to a¨ect the actions of people. ``Man observes that nature is sometimes

2

Chapter 1

Figure 1.1 Magical formula to exorcise Shabriri.

hostile and sometimes favourable to him and in this fact he sees a manifestation of the anger of the spiritsÐor of their good disposition. One must, then, treat the spirits as one would treat men whom one cannot dispense with'' (Bouisson 1960, 95; emphasis mine). Because speech and music were known to in¯uence the actions of other people, sometimes almost miraculously, many people felt there was a strong basis for believing that they would in¯uence the ``spirits,'' the forces that generated the natural phenomena that ruled their lives. An example of an incantation is the formula (®g. 1.1) from the Babylonian Talmud used to exorcise the demon Shabriri, who was said to cause blindness. The magical formula, repeated before drinking at night, consists of the demon's name (written here in Hebrew letters on the top line of the ®gure), pronounced six times, with ever fewer of its syllables (Shabriri, Briri, Riri, Iri, Ri, I). The demon is supposed to waste away as the syllables of his name are removed and thus not be able to do his evil work of blinding unsuspecting drinkers from the water jar. As illustrated by the Shabriri formula, incantational magic involves formulas of rhythms, intonations, and words performed in the correct order. Of course order is important even in ordinary speech. Indeed, it is a truism that word order can determine meaning, for example, ``sharkeating man'' versus ``man-eating shark.'' And intonation can also a¨ect meaning. For example, rising pitch toward the end of an English sen-

Magic, Ritual, and Dynamics

3

tence indicates a question; falling pitch indicates a statement of fact. Listen to yourself pronounce the following sentences: ®rst, ``The eggs are done?'' and then, ``The eggs are done.'' In poetry, the e¨ect of word order, rhythm, and intonation is even more striking. Read the following poem, by Robert Bringhurst (1995, 100) out loud: II Parable of the Harps In the drum of the heart Are the hoofbeats of horsesÐthe horse Of the muscles, the horse of the bones. In the ¯utes of the bones are the voices Of ®shesÐthe ®sh of the belly, The ®sh of the ®ngers and limbs. In the streams of the limbs We are swimming with ®shes And fording with lathering horses. Love, in this bed full of horses And ®shes, I bring to the resonant gourds Of your breasts the harps of my hands.

You can almost feel the horses thundering in your heart, or feel yourself in the river, swimming ¯uidly with the ®shes. But try reading aloud the following permutation of the poem: the swimming In bones. Are the Love, of the horsesÐthe horse the muscles, the horse Of the breasts heart In the bones I bring of the ®sh Of ®shesÐthe lathering of belly, fording The resonant gourds. ¯utes of the harps In the drum of ®shes limbs We are And ®shes, are the streams with limbs And ®sh with horses. in this of the bed of ®ngers to the horses and of the voices Of your hoofbeats the full of my hands.

Not only has the meaning of the poem been destroyed, but also the meaning of its rhythm and intonation. Indeed, it is no longer a poem, only a word salad. A good poem can bring tears to the eyes. Word salad just confuses, and can even signal mental illness. For example, people with untreated schizophrenia often utter strings of words that resemble the ``tossed'' version of Bringhurst's poem.

4

1.2

Chapter 1

Dynamics

The unfolding of behavior in time, especially that of human speech, has been a dominant theme in our understanding and subjugation of nature, as re¯ected by the dominant role played by dynamics in science, especially in physics, from Aristotle to Galileo to Kepler to Newton to Einstein and into the modern era. Quantum electrodynamics (QED), the foundation stone of modern physics, is essentially dynamical; indeed, according to Feynman (1985), it consists of only ``three little actions'' (emphasis mine): electrons and photons move in space-time and electrons scatter photons. Given the historical importance of magic, music, and rhythm, the preeminent role of dynamics in physics, and the emphasis on change over time in the other sciences (e.g., reactions in chemistry and evolution and development in biology), it is somewhat surprising that the dominant approaches to cognitive science, and to psychology as a whole, are statical rather than dynamical. That is, empirical laws and theoretical statements, when expressed formally, are written as B ˆ f …x1 ; x2 ; x3 ; . . . ; xn †;

…1:1†

where time is not a relevant variable, rather than as B ˆ f …x1 ; x2 ; x3 ; . . . ; xn ; t†;

…1:2†

where time enters as an important variable. A good example of a statical law in psychology is the well-known psychophysical power law, sometimes called ``Stevens's law,'' expressed in the following equation: R ˆ cS m ;

…1:3†

which describes reasonably well how responses (R) in a psychophysical scaling experiment vary with the intensity of the stimulus (S). For example, we might ask a subject to give a number (a ``magnitude estimation'') that indicates how loud a sound seems each time it occurs. When we play several sounds, each of a di¨erent intensity, several times each, and plot our subjects' average responses on a graph (see ®g. 1.2), the results closely ®t the curve of equation 1.3, which we can then use to calculate a value of R close to the one our subjects would give for any sound, even those to which they did not give a number. Because di¨erent people give similar numbers to the same sounds (Stevens 1975), this procedure is used routinely to predict people's responses to sounds when designing music halls, airports, and hearing aids.

Magic, Ritual, and Dynamics

5

Figure 1.2 Stevens's law (curve) summarizes psychophysical scaling data (dots).

As useful as Stevens's law is, it does not describe everything of interest about subjects' responses in a psychophysical scaling experiment. Subjects do not always give the same numbers when presented with the ``same'' physical stimulus (the error bars were omitted in ®g. 1.2). Of course, as the quotation marks around ``same'' indicate, all stimuli di¨er, no matter how hard we try to get them to be the same. Air currents, voltage ¯uctuations in a sound generator, increasing fatigue of loudspeaker cones, rising or falling temperature and air pressure, and so forth together produce trial-to-trial ¯uctuations in sound pressure at the eardrum even for the same nominal sound stimulus. And subjects change, too. Blood pressure rises and falls, attention wavers, digestion proceeds, viruses and bacteria multiply and die, and so forth. It would be a miracle if subjects and stimuli were exactly the same on any two occasions. In fact, subjects often give dramatically di¨erent responses to the ``same'' stimulus at di¨erent times, and often give the same response to dramatically di¨erent stimuli presented on di¨erent occasions. A psychophysical scaling experiment is necessarily extended in time, and the behavior involved in making psychophysical judgments ¯uctuates over time, even when the experimenter strives to make the situation identical from moment to moment. How should we deal with these ¯uctuations? In psychophysics, one tradition, attributed to S. S. Stevens (1975), has been to ignore them. The justi®cation is that we are not usually interested in the ¯uctuations in subjects' blood pressure, digestion, sleepiness, and so forth over the

6

Chapter 1

course of a psychophysical scaling experiment. Nor are we interested in the unavoidable di¨erences in the stimulus magnitude from trial to trial. These e¨ects re¯ect our incomplete control over the experimental situation and are nuisances. We therefore agree simply to call these ¯uctuations ``error variance,'' to be ``averaged away'' (as in ®g. 1.2). A di¨erent tradition, beginning with Fechner (1860) and continuing through Thurstone (1927) and Green and Swets (1966), is to use the amount (but not the timing) of the ¯uctuations to measure sensation and sensory discrimination. Thus Fechner built his famous law of sensation on what we now call ``Weber's law,'' that the di¨erence threshold (the smallest intensity di¨erence that can be reliably detected) is proportional to the stimulus intensity at which it is measured, DI ˆ kI , where DI is the di¨erence threshold and I the stimulus intensity. This proportionality arises because response variability generally increases with stimulus intensity. In signal detection theory (and related approaches), an important measure of stimulus discriminability, d 0 , is de®ned as the di¨erence between the means of two assumed probability distributions of sensory e¨ects of stimuli divided by their common standard deviation, the latter representing the amount of ¯uctuation in those sensory e¨ects. Just how those ¯uctuations are distributed over time is irrelevant (but see Link 1994). Thus, whether a nuisance or a fundamental concept, psychophysical variability is usually treated as ``error'' rather than as temporally distributed information about cognitive processes. Both traditions ignore time, even though ¯uctuations over time can provide fundamental information about the processes generating the behavior in question. This point applies not only to psychophysics, but also to most of the work being done in cognitive science today. It is undeniable that in some cases the temporal distribution of response variability adds nothing to our understanding of the phenomenon we are studying. On the other hand, however uninteresting the temporal distribution of these ¯uctuations might seem to be, it takes on new meaning when juxtaposed with several signi®cant observations from physics and biology and from physical methods applied to human behaviors. First, such temporally distributed ¯uctuations are ubiquitous; indeed, in quantum mechanics and statistical physics, they are fundamental (see chap. 13). Second, these random processes can actually ``drive'' physical and biological phenomena by falling into one or the other of several ``basins of attraction,'' depending on random ¯uctuations early in the process. They can determine, for example, which of two equally rich food sources will be exploited by an ant colony (e.g., Beckers et al. 1990). Random

Magic, Ritual, and Dynamics

7

choices made by the ®rst ants to reach a choice point bias the probabilities of choices made by ants reaching the choice point at a later time (via the pheromones they deposit), leading to a strong ``preference'' for one food source over the other. Importantly, random choice ¯uctuations of the same size occurring at a later time in the process, once a preference has developed, have no e¨ect on which source is preferred. In this and in many other cases, the timing of the ¯uctuations is all important. Third, the random ¯uctuations can convey valuable information when observed in human behavior. For example, ¯uctuations that occur with a certain type of periodicity, called ``1=f '' or ``pink'' noise (see chap. 15), are diagnostic of stochastic processes that interact at several scales of time; 1=f noise characterizes the residual response variability in standard experimental tasks of cognitive science (see chap. 16). Finally, even deterministic processes can create ¯uctuations that appear random (Chan and Tong 2001). For example, the logistic di¨erence equation Yi ˆ aYiÿ1 …1 ÿ Yiÿ1 †;

…1:4†

although simple and completely deterministic, demonstrates extremely complicated, e¨ectively random behavior for values of a > 3:58. Only dynamical analysis of time series of behaviors can reveal the di¨erences between such deterministic processes and processes characterized by white noise. In making these same points, other authors (e.g., Gregson 1983, 1988; Port and van Gelder 1995; Kelso 1995) have also argued that dynamics should be central to cognitive science and to psychology as a whole. Nevertheless, even though more and more cognitive scientists are using dynamical metaphors, and even though some are using dynamical system theory to analyze behavioral experiments, the practice of dynamics is still not widespread. This is so partly because dynamics is complicated and di½cult, requiring considerable sophistication in mathematics and other disciplines with a large technical content, and partly because some practitioners of the dynamical approach (e.g., Gregson 1983, 1988) have set it against the more traditional approaches, arguing for a Kuhnian paradigm shift in psychology. Unfortunately for the latter, not all statical theories of psychology can be subsumed as special cases under dynamical theories (see chaps. 8 and 9). Finally, psychology's reluctance to embrace dynamics lies partly in our having few well-worked-out examples of the bene®ts that accrue when the dynamical approach is taken (see, for example, Kelso 1995). In a discipline such as economics, where prediction of the next values of a time series (e.g., stock market prices) has obvious bene®ts,

8

Chapter 1

dynamics is understandably central. In psychology, however, the bene®ts of dynamical analysis are not so clear. Sometimes a dynamical analysis simply generalizes a statical theory to other situations and data, particularly where values ¯uctuate in time, as in the psychophysical case (see chap. 11). Sometimes, however, the dynamical analysis provides a fundamentally di¨erent theoretical approach to an old problem, with novel predictions following from the dynamical theory. Beer (2000) describes one such dramatic case in the ``A-not-B'' error in infant reaching: infants 7 to 12 months old continue to reach for an object in the one of two opaque containers they have been trained to reach for, even after they have seen the object being hidden in the other one. Whereas Piaget argued that the error is the result of applying an immature concept of object permanence to the task, a dynamical analysis suggests that it is caused by an immature goal-directed reaching system (Thelen et al. in press). The dynamical model also accounts for the dramatic context e¨ects that have been observed in this task. Moreover, it makes the novel prediction that, under the right conditions, the error should be observed in older children as well because it arises from general properties of the perceptual-cognitive-motor system controlling goal-directed reaching. Thus, if con®rmed, the dynamical theory would dramatically change our understanding of infants' performance in goaldirected reaching, and perhaps even our assessment of concepts such as object permanence in child development. In light of the possible bene®ts, this book aims to overcome the considerable obstacles to the centrality of dynamical analysis in cognitive science. It aims to describe some tools of dynamical analysis simply and clearly, so that even mathematically unsophisticated researchers can see how they are used. It aims to show how dynamical and statical approaches are complementary and mutually informative. And ®nally, it aims to provide examples of the increases in understanding that accrue when cognitive science is informed by both dynamical and statical analyses.

Chapter 2 Sequence

The various sections of the Torah were not given in their correct order. For if they had . . . anyone who read them would be able to wake the dead and perform miracles. For this reason the correct order of the Torah [is] known only to the Holy One. . . . ÐRabbi Eleazar ben Pedath, quoted in Gershom Gerhard Scholem, On the Kabbalah and Its Symbolism The mind which any human science can describe can never be an adequate representation of the mind which can make that science. And the process of correcting that inadequacy must follow the serial steps of an in®nite regress. ÐJ. W. Dunne, The Serial Universe Just about all of the interesting hypotheses we have about how social systems function imply at their base an imagined scenario of interaction, a scenario invariably sequential in character. ÐJohn M. Gottman and A. K. Roy, Sequential Analysis

2.1

The Serial Universe

It is often best to begin at the beginning. In the case of sequence, the beginning has to do with the nature of the universe itself. In his charming little book, from which the second epigraph above is taken, Dunne (1934, 28) argued that the universe ``as it appears to human science must needs be an in®nite regress'' (emphasis his). In this context, a ``regress'' is a question that can be answered only by asking another, similar question, which of course leads to asking still another, similar question, and so forth, to in®nity. For example, a child learning arithmetic might attempt to answer the question ``What is the largest integer?'' The child ®rst considers whether there is a larger number than ``1'' and discovers that ``2'' is larger than ``1,'' which leads to another question, whether there is a number larger than ``2,'' and so forth. Because, ``1'' can always be added to any

10

Chapter 2

Figure 2.1 Dunne's serial painter (1934).

integer, this series of questions can go on forever. Although Dunn felt we cannot know whether the universe is actually an in®nite regress, he tried to convince his readers that, as revealed by science, it must be so. Without going into the details of Dunne's arguments, let us brie¯y consider his theory of ``serialism,'' which contains a compelling case for treating the objects of science, including cognitive science, as observations in time. Figure 2.1, from Dunne (1934), illustrates the dilemma in which we ®nd ourselves. We are like the artist trying to portray himself and all of his knowledge of the scene before him in a single painting. He is frustrated in this by the fact that he himself is always outside the frame of the picture he paints. Thus he must paint a series of pictures, each perhaps more complete than the last, but all subject to the limitation that the information displayed must be less than the information involved in creating the display. GoÈdel (1931) dealt with a similar problem, how to prove all of the true statements within a consistent formal system, and Penrose (1994) argued that the resulting nonalgorithmic nature of human understanding of the world is the hallmark of human consciousness.

Sequence

11

Dunne's theory of serialism (1934) contends that, as self-conscious beings, we are able to di¨erentiate ourselves from ``the world'' (including other humans). Like the artist pictured above, however, we are unable to treat this predicament systematically except as an in®nite regress. And because time is regressive in just this sense, each instant outside of and containing all of the prior instants, viewing all experience in terms of time provides the required descriptive framework. This framework introduces the notion of change and allows us to consider causing changes both in ourselves and in the world. Change, in turn, creates the possibility of experimental science: it allows us to interfere with what we observe and to record the results of that interference. This interpretation of experiments is an axiom of quantum physics. Thus, at the most fundamental level of science, we are forced to be regressive, to describe events in time, and to consider sequence. 2.2

The Problem of Serial Order in Behavior

Karl Lashley is probably best known today as the psychologist who searched for, but failed to ®nd, the engram (the speci®c, localized neural substrate upon which memories were laid down) in a series of brilliant experiments with rats (e.g., Lashley 1950). However, he was also one of the earliest psychologists to point out (Lashley 1951) that every type of behavior, from moving to speaking to designing a building or writing a book, temporally integrates sequences of actions in complex waysÐways that cannot be explained by how the behaving organism responds to sequences of external stimuli. He asserted that especially complex action sequences are absent in animals that lack a cerebral cortex, such as insects. He considered complex, temporally integrated actions to be especially characteristic of human behavior. He rejected the ``associative chain theory,'' in which execution of each element in a chain of re¯exes provides the stimulus for the next, as an explanation for temporal integration because it is unable to account for myriad empirical facts. One striking example is in typing on a keyboard, in which the various motor acts can follow each other in any progression, such as the various ones that make up the very sentences I am typing here. The letter ``e'' does not invariably follow the letter ``r,'' the letter ``a,'' or any other letter, although in English, the letter ``z'' seldom (and then only in typos) follows the letter ``c.'' (in Polish, on the other hand, . . . .). Similarly, the phonemes that make up a spoken word can follow each otherÐand coarticulateÐin nearly any sequence at all. Moreover, sequences never before uttered by a

12

Chapter 2

person (take the novel pseudoword ``ba½t'') can be uttered perfectly the ®rst time, without any learning whatsoever. In speaking, typing, running in unfamiliar terrain, sight-reading music, and so forth, the novel action sequence consists of low-level units organized into a hierarchical structure that seems to be governed from above. Lashley (1951) rejected ``commonsense'' answers about how temporal order is imposed on the elementary units of an action sequence, arguing for a new approach, one in which the brain is a dynamic, constantly active system. An input to such a system, rather than exciting a single, isolated, re¯exive pathway, causes widespread changes in the entire brain system, much as gusts of wind, tossed stones, a swimmer's strokes, and so forth cause ripples on the surface of a lake. And just as the surface of the water is always active and its activity always the product of the many competing and cooperating in¯uences, in Lashley's view, action is temporally integrated under the in¯uence of reverberatory circuits in the brain created by the interaction of inputs with ongoing and preceding activity. More speci®cally, Lashley contended that all stimuli perceived and all actions generated by an organism are distinguished in terms of a system of spatial coordinates; that, once representations (``traces'') of stimuli or movements have been discriminated in this system, they are ``scanned'' by a separate system of the brain that translates the spatial into a temporal order. Similarly, a series of stimuli received in a particular temporal order are represented internally as di¨erent loci in the spatial system, and then read out in the appropriate (or a changed) temporal order by the scanning system. Finally, a prescribed series of actions, naturally represented in di¨erent spatial loci within the motor system, is activated by the scanning system and thus produced externally in the appropriate temporal sequence. Lamenting that Lashley had not described his scanning mechanism in su½cient detail, Bryden (1967) proposed another, slightly more detailed mechanism, in which stored ``ordering systems'' specify inhibitory and excitatory interactions between memory traces of stimuli and action units in their control of output pathways. The overlearned order of reading words from left to right in English, for example, imposes a di½cult-to-change ordering on human subjects' reporting of visual and auditory stimuli. Bryden also used his theory to explain both the order in which items to be remembered are rehearsed, and data in several other paradigms. It is remarkable how modern Lashley's words sound when read today, even though his program is only beginning to be followed. An excellent example of recent work in the spirit of Lashley's dynamical approach is

Sequence

13

Clancey's detailed and fascinating computational analysis (1999) of what he calls ``conceptual coordination.'' Whereas, in ``situated cognition'' (Clancey 1997), concepts could be coupled (simultaneously activated) and inferred (sequentially activated) under the control of a (physical, social, mental) situation, in conceptual coordination, associations between concepts in memory operate as either simultaneous or successive neural activations in the brain. Clancey (1999) also argued that many of the organizers of cognitive behavior thought to be innate, such as grammar, might be learned and might constitute a kind of virtual machine running in the brain's basic neural architecture (cf. chap. 35). In Clancey's system (1999, xvi), ``ordered behavior in time might be constructed and composed dynamically from simple processes of sequential and synchronous activation of categories.'' This view contrasts mightily with the usual analyses of memory and conceptualization, which emphasize static, connected structures such as semantic networks, and provides a dramatic example of the conceptual bene®ts that accrue from a dynamical analysis of cognition. 2.3

Markovian Analysis of Behavior

One of the most direct ways to analyze sequences of behavior, at least when the behavior can be categorized into several simple categories, is to consider the sequence as a series of transitions between states of a behaving system, where each transition occurs with a particular probability. When, in a sequence of probabilistic state transitions, the next state depends only on the current state, we have what is called a ``Markov chain,'' and the process from which the chain arises is called a ``Markov process'' (see also chap. 12). A Markov process is described fundamentally by its initial state vector, which gives the probabilities for each of the states of the system at the start of a chain, and its transition matrix, which summarizes the probabilities of transitions between each of the states. An example of a simple transition matrix for a threestate Markov process is given in ®gure 2.2. Notice that, because the probabilities in the matrix are conditional on the state at time t, for example, P(0 at t ‡ 1 given 1 at t) ˆ P…0 j 1† ˆ 0:1, the rows in ®gure 2.2 sum to 1.0. Notice, also, that a transition from states 0 and 1 to any of the three states is possible; for example, the probability of a 0 ! 1 transition is 0.6. However, if the chain arrives at state 2 the only transition possible is to state 0. A series of states, or a Markov chain, generated by this transition matrix would look something like this:

14

Chapter 2

Figure 2.2 Typical Markov transition matrix.

0 1 1 1 1 1 1 2 0 0 1 0 1 1 1 1 1 0 1 1 1 1 0 0 2 0 1 2 0 1 1 1 1 1 1 1 . . . Notice that this chain fragment is dominated by state 1, the most likely transition out of both state 0 and state 1; once in state 1, the process either stays there (with probability 0.8), goes back to state 0, or goes to state 2, from which it always goes back to state 0 again. State 1 is close to being an absorbing state (which it would actually be if the probability of a 1 ! 1 transition were 1.0). Occasionally the lower probability transitions 0 ! 0 or 0 ! 2 occur, both of which result in 0 again and another chance to enter state 1 with high probability. Interpreting the states as categories of behavior, say reading the TV Guide, watching a TV show, and getting a snack for states 0, 1, and 2, respectively, we could use the transition matrix in ®gure 2.2 as a model for the sequential behavior of a person who likes to watch TV. Notice that ®gure 2.2 does not exactly describe any one chain, only the overall ``look'' of all chains generated by sampling the probability distributions (``realizations'') described by the transition matrix. The fundamental mathematics of this approach was developed extensively by mathematicians and physicists, and the approach has been applied in areas of psychology concerned with change ever since Miller 1952 introduced the idea to us. The mathematics is closely related to the analysis of contingency tables with chi-square statistics, and also to Shannon and Weaver's information theory (1949). Indeed, a type of chisquare test can be used to establish the necessity of a Markov analysis based on observation of a particular sequence of behaviors. Other statistics can also be calculated that allow comparison of transition matrices across individuals or groups. In the 1950s, Markov modeling was applied especially successfully to animal learning (e.g., Bush and Mosteller 1955) and, in the 1960s, to human concept learning (Trabasso and Bower 1968). In the latter application, the states involved were explicitly cognitive states that represented the subject's internal hypothesis as to the concept the experimenter was using the generate the series of exemplars and nonexemplars to which the subject had to respond. Because Trabasso and

Sequence

15

Figure 2.3 Frequency and probability matrices for the talking sequence.

Bower's analysis (1968) is somewhat complicated, however, I will illustrate the Markov approach with an example from the extensive work on sequential analysis of social interactions described in Gottman and Roy 1990, which also described in detail how to use the technique for analysis of a wide range of behaviors, including language behavior (cf. sect. 2.2). Because the techniques are easy to use, the statistics familiar, and the mathematics related to very deep results in physics (see chap. 13), sequential analysis via Markov models remains one of the most attractive tools for dynamical cognitive science. As an example of the ®rst steps in a Markov analysis of a sequence of behaviors, consider the following sequence of states described by Gottman and Roy (1990) from a study of conversational interaction by Ja¨e and Feldstein (1970). Here the states refer to whether a particular one of the conversationalists is talking (state 1) or remaining silent (state 0Ðthe other person is talking) and describe one minute of conversation: 111111111000111111111111111111111111111111111111111111111 111111111111111111111000000000000000001111111111111111111 111111111111111001111111111111111110011111111111111111111 1111111111110100000001111111 The frequency and probability transition matrices for this sequence are shown in ®gure 2.3. The frequency matrix was created by counting the numbers of transitions of each type in the sequence above; the probability matrix (whose entries actually only estimate the ``true'' probabilities), by dividing each entry in the frequency matrix by its row sum (e.g., 25=31 ˆ 0:81). The value of w 2 for this frequency matrix is 118.99, which is much larger than the 99.9th percentile of the w 2 distribution with 1 degree of freedom (10.82), indicating that the observed frequencies of

16

Chapter 2

transitions are very unlikely ( p f 0:001) to have been generated by sequentially independent samples from the overall, unconditional probability distributions of the two states. In other words, there is tendency for the speaker, once having started talking (in state 1) to persist in talking, and a similar, although not quite as strong, tendency to persist in silence once having fallen silent. Another way to see this is to calculate the unconditional probabilities of silence and talking by dividing the total numbers of occurrences of the two states (31 and 168 for silence and talking) by the total number of elements in the chain (199), yielding P(silent) ˆ 0:16 and P(talking) ˆ 0:84. These probabilities can be compared with the conditional probabilities from the transition matrix, P…0 j 0† ˆ 0:81 and P…1 j 1† ˆ 0:96. Clearly, because these conditional probabilities are larger than the corresponding unconditional probabilities, there is a tendency to remain in the same state over time, even when the transition that brought one to that state is unlikely, for example, P…0 j 1† ˆ 0:04. The Markov transition matrix captures this tendency simply and elegantly, and predicts similar, although not identical, sequences in other settings. Similar analyses (e.g., Gottman and Roy 1990) have been e¨ectively applied to marital and other social interactions, resulting in useful predictions and insight about underlying processes.

Chapter 3 Rhythms of Behavior

It can now be said with assurance that individuals are dominated in their behavior by complex hierarchies of interlocking rhythms. Furthermore, these same interlocking rhythms are comparable to fundamental themes in a symphonic score, a keystone in the interpersonal processes between mates, co-workers, and organizations of all types on the interpersonal level as well as across cultural boundaries. I am convinced that it will ultimately be proved that almost every facet of human behavior is involved in the rhythmic process. ÐEdward Twitchell Hall, The Dance of Life

3.1

The Dance of Life

Hall (1983) was one of the ®rst to observe that humans in all cultures are engaged in a rhythmical ``dance'' when they interact with each other. Studying ®lms of people interacting in a wide range of situations, from laboratory to ®eld, and both within and across cultures, he documented both the actual rhythms and the intriguing variety of these rhythms, especially across cultures. Moreover, he also described the dramatic ways in which rhythms can a¨ect performance, relationships, enjoyment of life, and the outcomes of interactions between cultures. Rhythm is one of the most elementary and at the same time one of the most profound categories of sequential behavior. Rhythmical time series are simple and yet they exhibit a compelling structure. Interestingly, the rhythms of interpersonal behavior are not easy to see or to describe in English. When Hall asked some of his Chicago-born American students to view his ®lms of people interacting and to report what they saw, it usually took them many hours of viewing before they began to notice and interpret the nonverbal gestures and other movements of the people in the ®lm. On the other hand, when he asked SpanishAmerican students to view the same ®lms, they were able to see the rhythmical dance right away. What was the di¨erence? According

18

Chapter 3

to Hall, the Spanish-American students came from a culture in which rhythm and repetition are foundations of interaction, whereas the other students were from a verbal culture that trained people to seek variety and avoid repetition. Moreover, the Latin culture (and others, such as African) emphasize the connectedness of people and their embeddedness in a cultural context, whereas the dominant North American culture emphasizes the uniqueness of individuals and their isolation from others. This leads also to di¨erent views of time in those cultures: the Latin and African cultures have a polychronic, or ``many clocks'' view of time, whereas the North American culture has a monochronic view of time, with just one clock, the physical one on the wall or on the wrist that says when events such as classes, meetings, or vacations are to begin and end. For Hall, the rhythms of life generate a great variety of types of time, illustrated in his mandala of time in ®gure 3.1. Although we will not explore most of these types of time, their characterization is consistent with the view of time as constructed and evolving put forth in chapter 4. It is instructive to com-

Figure 3.1 Mandala of time. (After Hall 1983.)

Rhythms of Behavior

19

pare this polychronic view of time with the nested hierarchy of times proposed by Fraser (1982), and discussed in chapter 4. 3.2

Music and Rhythm

One topic that creates both a desire to attend to rhythm and a vocabulary with which to describe it is music. Even here, however, we ®nd that Western music and music in the rest of the world treat rhythm quite differently. In his beautiful book, Robert Jourdain (1997) explores the many facets of music and its construction and appreciation by the human brain. He considers two very di¨erent notions of rhythm that dovetail nicely with Hall's conceptions. Meter consists of metronomic patterns of accentuated beats, whereas phrasing consists of more organic, irregular but structured temporal organizations of musical ``shapes.'' Or, as Jourdain (p. 130) puts it: ``Meter is brick, phrasing is poured concrete.'' The regular beat of rock music, indeed of most Western music, is meter at its strongest. Try tapping your ®ngers on a surface at a steady rate. Soon you will ®nd yourself accentuating some taps and not others: one-two, one-two, and so forth. Even perfect, metronomic pulses are organized into groups of beats (e.g., Povel 1981). The one-two-three, one-two-three meter of a Viennese waltz is what lays the foundation of the dance of the same name, a three-step that, at least for beginners, proceeds with one foot movement per beat. In parallel with the monochronic Western dance of life, most Western music is informed by a single meter, a monorhythm. In contrast, meter in ``world music'' is often polyrhythmic, consisting of several interacting meters. African drumming is perhaps the quintessential example of polyrhythm, often incorporating four or more di¨erent meters into a single fabric that seems to ebb and ¯ow like a tide. Such rhythmic elaboration is attractive when only a few, rather cumbersome instruments, like drums, are available. As an example of the powerful in¯uence of phrasing on our appreciation of music, Jourdain analyses ``The Pink Panther'' by Henry Mancini. Figure 3.2 shows the score for the ®rst few bars of this unforgettable piece of music. The rather simple meter is shown at the bottom of the ®gure; the groups of notes comprising musical phrases and subphrases are indicated at the top. Note that the phrases are easier to understand when listening to or humming the music than when visually inspecting the score. You might wish to try humming the tune from memory while perusing ®gure 3.2. It is clear that the phrasing builds a complex hierarchical structure. This structure strongly organizes our perceptions of the music and, in this

20

Chapter 3

Figure 3.2 Meter and phrasing in ``The Pink Panther.'' (From Jourdain 1997, p. 123. ( 1997 by Robert Jourdain. Reprinted by permission of HarperCollins Publishers, Inc.)

instance, creates even more complexity by sometimes accentuating and sometimes undercutting the meter. All of this creates a driving piece of music that we cannot help moving to, although these movements do not include any kind of simple underlying oscillations, such as those made when we tap out the beat of a piece of rock music. Meter is clocklike; it is involved in measuring temporal intervals. It creates temporal order with an underlying pulse, which entrains the other meters, even complex ones like syncopation and polyrhythms. In contrast, the perceptual present, memory, and anticipation work together to generate phrasing, or the large-scale temporal structure of music. In dance, we move to the rhythm of a piece, trying to feel or to convey the music's rhythmical structure, both meter and phrasing. Jourdain argues, however, that rhythm does not arise from movement. Rather, it is generated by the brain in an e¨ort to organize perception and cognition along the temporal dimension. Rhythm, especially phrasing, is largely learned and may depend on oscillators in the brain that are engaged in structuring experience in ``phrases'' (see chap. 31). Some phrasings are clearly related to the organic way we act in the world, and to the temporal contingencies of the evolutionary niche we occupy. Others are constructed in creative acts of elaboration of our musical vocabulary. 3.3

Rhythms in the Brain

In another beautiful book Scott Kelso (1995, xvii), describes how the brain, conceptualized as a ``pattern-forming, self-organized, dynamical system poised on the brink of instability,'' creates rhythms. He begins with an account of a clever ®nger-twiddling experiment, which he explains

Rhythms of Behavior

21

using the Haken, Kelso, and Bunz (HKB; 1985) model of coordination. An example of synergetics, the interdisciplinary theory of pattern formation created by Haken (1983), the HKB model is the basis for Kelso's description of how the brain organizes itself to create the rhythmical ¯ow of behavior, including perception and learning. Although Kelso used a back-and-forth motion, let us try a tapping version of Kelso's task, one that connects more closely with musical rhythm. Place your hands on a surface in front of you, thumbs touching, index ®ngers in the air pointing away from you, and the other three ®ngers of each hand curled underneath, out of the way. Now try tapping one ®nger on the surface while lifting the other one in the opposite direction, away from the surface, then lifting the tapping ®nger and tapping with the other one, repeating this alternate tapping motion at the same, relatively slow rate for a while (imagine you are playing a bongo drum for a very slow Latin ballad). This tapping is called ``antiphase movement'' because the homologous muscle groups in your two hands are doing di¨erent, opposing, things at the same time. Now try ``in-phase'' movements, where your ®ngers are moving in the same direction (up or down) and tapping at the same moment. Here homologous muscles that move your index ®ngers are doing the same thing at the same moments. Now go back to antiphase tapping at a slow rate and begin to speed up your tapping rate (following a metronome's clicks is a good way to force the higher tapping rate). As the tapping rate increases, you will ®nd it increasingly di½cult to continue tapping in antiphase, ®nally succumbing and switching to inphase tapping. Now try beginning with in-phase tapping. You will ®nd that you are comfortable tapping in phase at any rate, including the very slow ones at which antiphase tapping also seems comfortable. We could say that in-phase tapping is ``stable'' over the entire range of tapping rates, whereas antiphase tapping is only stable at slower rates, becoming ``unstable'' at higher rates. The instability causes the switch to in-phase tapping. The HKB model, brie¯y described below, explains this dramatic di¨erence in the stability of the two tapping patterns (for the full account, see Kelso 1995). We can describe the results of Kelso's experiment (1995) in terms of a single variable, the relative phase, f, between the movements of the two ®ngers, measured by subtracting where one ®nger is in its up-down cycle (its phase) from where the other one is at the same moment in time, relative to a canonical position (say touching the surface). If we consider an entire up-down cycle to be a circle, with the canonical position at 0, each cycle has magnitude 2p (the number of radians in a circle;

22

Chapter 3

p ˆ 3:141593 . . .); thus relative phase can be anything between 0 and 2p. For example, if one ®nger is down (phase ˆ 0, touching the table) when the other is up (phase ˆ p, halfway around the circle), then the relative phase is p ÿ 0 ˆ p (or vice versa, 0 ÿ p ˆ ÿp). The ®nger-tapping experiment result is that a relative phase of 0 (or, equivalently, 2p, in phase), with the ®ngers always at the same phase of their movements, is stable at all rates, whereas a relative phase of p (or ÿp, antiphase), with each ®nger moving in the opposite direction from the other, is only stable at slow rates, switching to f ˆ 0 at some critical speed of cycling (degrees per second). Notice that relative phase is unchanged if we reverse which ®nger is in which position. For example, f ˆ p, whether the left ®nger is down and the right one up or vice versa. Relative phase is also unchanged if we shift time by one entire period, adding or subtracting 2p. For example, 0 and 2p are identical, as are relative phases of p and 3p (or ÿp and p). We can express these spatial and temporal symmetries in terms of a mathematical function, V …f†, that stands for the behavior of the two ®ngers: V …f† ˆ V …ÿf† and V …f ‡ 2p† ˆ V …f†, respectively. What is the simplest function V that could display these two symmetries and also describe the results of the experiment? As Kelso (1995) explains, temporal symmetry implies that V is periodic and thus can be expressed as a Fourier series, a sum of sine and cosine functions (see, for example, Ramirez 1985 and chap. 14 for more). Spatial symmetry means that V has only cosine functions because cos…x† ˆ cos…ÿx†. Because we want the simplest model that just does the job, we need only the ®rst two terms in the series: V …f† ˆ ÿa cos f ÿ b cos 2f;

…3:1†

the equation for the HKB model of coordination. Figure 3.3 shows graphs of equation 3.1 for various values of b=a, which is interpreted as related to rate of tapping, f , by b=a ˆ k=f , where k is a constant. Large values of b=a correspond to slow tapping, and the smaller the value of b=a, the faster the tapping. Finally, for any value of a, if b ˆ 0, b=a ˆ 0, and tapping is the fastest it can be ( f approaches in®nity). When this happens, V …f† ˆ ÿa cos f, as shown in the bottom right-hand graph. We can interpret the graphs in ®gure 3.3 as a two-dimensional landscape on which a ball (representing a particular value of the phase relation f) is rolling around, subject to a downward force (e.g., gravity). When the ball is in a valley, it cannot get out unless energy is exerted to raise it above one of the adjacent hills (a ``potential barrier''), which represents a stable state of the system (black ball). If there is no adjacent hill, the ball could easily roll downhill into a valley, and it is in an un-

Rhythms of Behavior

23

Figure 3.3 The Haken, Kelso, and Bunz (HKB; 1985) model of coordination. (From Kelso 1995. Reprinted by permission of The MIT Press.)

stable state (white ball). Notice that there is always a center valley with a black ball in it, regardless of the value of b=a. Thus, if tapping in phase (ball in center valley, f ˆ 0), a person can easily keep on doing this, no matter whether the tapping rate becomes higher (b=a gets smaller) or lower (b=a gets larger). However, if a person begins tapping in antiphase (f ˆ p or f ˆ ÿp) at a high rate (bottom row of graphs), the state is unstable, and the slightest ¯uctuation will send the ball rolling downhill toward the nearest (and only) stable state, f ˆ 0. If a person begins tapping at a slow rate, say at b=a ˆ 1, and gradually speeds up, the terrain gradually changes with the tapping rate until, once again, the system becomes unstable and eventually falls into the stable state, f ˆ 0. This happens at some critical value of b=a, in this case near b=a ˆ 0:25. The parameters b and a, and the critical value of b=a, could be di¨erent for each person, but when the correct values are ®tted to the data, the model

24

Chapter 3

describes each person's phase-switching behavior quite nicely. Moreover, because the ®ngers are controlled by the brain, the HKB model can be interpreted as a description of the self-organized, coordinative behavior of those parts of the brain involved in generating rhythmical movements, just the kind of thing demanded by both music theory and by the dance of life. The HKB model is an excellent example of the application of dynamical systems theory to explain the temporal integration of behavior. The central role played by ¯uctuations in the dynamics of behavior, speci®cally, in precipitating movement between states and, more generally, in synergetics, is a major theme of the chapters to come.

Chapter 4 Time

Time ¯ies, time dies. ÐJonathan Larson, ``One Song, Glory,'' from Rent, The Musical What really matters anyway is not how we de®ne time, but how we measure it. ÐRichard Phillips Feynman, Robert B. Leighton, and Matthew L. Sands, The Feynman Lectures on Physics

Ordering behaviors in a sequence is not enough. For many reasons that will become apparent if they are not already, we need to locate behaviors in time. But what is time? Despite its central role in their subject, most authors writing about dynamics do not discuss the nature of time, perhaps because they realize just how di½cult it is to discuss properly. Failure to deal with the nature of something as fundamental to dynamics as time, however, can lead to inadequacies in the resulting theory. Changes of a system over time are observed, categorized, and analyzed, but the role of time in these changes is relegated to a symbol (usually, t) in a di¨erential equation. (On the deep problems time poses, see Russell 1945 and Van Bendegem 1992 for reviews of work before and after 1945, respectively; see also Coren, Ward, and Enns 1999 for a conventional discussion of the role of time in perception and cognition.) Here, in touching on some of the very deep problems that remain, I adumbrate a coherent view that will serve our present purposes. 4.1

Space-Time

The time employed assumptively in most discussions of dynamical systems, including the few that exist on the dynamics of behavior or of cognition, is the absolute space-time of Newtonian, or classical, physics. In this system, dimensions, in which moving particles exist, are in®nite in

26

Chapter 4

Figure 4.1 A Feynman diagram of an electron (straight lines) moving in time and space and emitting a photon (wavy line) at a certain time and place.

extent and axiomatically or de®nitively given. For example, ®gure 4.1 shows a ``Feynman diagram,'' which Richard Feynman (1985) invented to keep track of the possible paths of photons and electrons in doing calculations in quantum electrodynamics. Similar diagrams are used in discussions of cosmology to illustrate the evolution of the universe and in many other areas of physics. In such diagrams, space and time are represented as two dimensions: space is relegated to one dimension, and time to the other. In psychological diagrams, the space dimension might be changed to a psychological variable, such as anxiety or the probability of a correct response to a short-term memory probe, but the time variable is unchanged. This absolute time is treated as a variable in an equation and is measured by physical clocks. It appears either as an explicit variable, for example, in d ˆ rt, where d is distance, r is rate of movement, and t is time; or as an implicit variable, for example, in F ˆ ma, where F is force, m is mass, and a is acceleration, the rate of change of the rate of change of position, d…dx=dt†=dt. The calculus is not crucial to understand acceleration at this point; I will indicate and explain it when it is. Where did the idea of absolute time come from? According to Feynman, Leighton, and Sands (1963), there was no precise way to measure time when philosophers (e.g., Plato and Aristotle) ®rst began to discuss it. Galileo ®rst measured time intervals in his inclined plane experiments by counting his heartbeats (pulse). Later, to gain more precision, he created

Time

27

better clocks based on pendulums. The drive to create better clocks has continued until the present day, when physical time is measured by the vibrations of the cesium-60 atom and other exotic oscillators. Having a good measuring system for time also allows us to measure other important variables. For example, distance can also be measured in terms of absolute time as long as speed is constant. One useful application of this idea is in radar: radio waves are bounced o¨ objects, whose distance, d, is calculated from d ˆ ct=2, where c is the speed of light and t is the time from emission of the radio wave by the emitter until its detection by the receiving attenna. In contrast to the absolute space and time of physics, the space and time of Leibniz (1714/1956) were de®ned by relations among an in®nite number of particles (he called them ``monads'') that had no extension in either space or time. Although various aspects of Leibniz's philosophy are problematic from the point of view of dynamics, especially the assumption that monads cannot interact, the idea of relational space and time itself is not. Consider a set of identical elements, or particles, in the abstract. To speak of one, rather than another, there must be some property by which the one can be discriminated from the other. The most elementary such property is location in space. If the two di¨er in location, then I can point to one or the other and say, ``That is the one I mean.'' In physics, elementary particles are of this nature: two photons or two electrons of the otherwise same quantum properties are interchangeable except for a di¨erence in spatial location. When I can discriminate two otherwise identical elements in this way, I can say that space is ``generated'' by the relationship between the two. The unit of space is the ``di¨erence'' between the elements that allows the discrimination. Time then arises from treating the two (or more) spatial locations in some order. Interval (or duration) is generated by movement (change) from the ®rst to the second, whereas period is generated by movement from the ®rst to the second and back again. In the simplest case, interval is generated by two successive pointing events (cf. Wittgenstein 1961). Elements that can be distinguished spatially but not temporally are concurrent, or simultaneous. In some general systems theories, space and time are ``constructed'' through the interactions of the system with its environment which causes the di¨erentiation of initially identical elements, ®rst in space and then in time (Alvarez de Lorenzana 2000). Whether we consider time to be relational or absolute, a fundamental relationship between space and time is that time can be represented by distribution in space plus order (or sequence). By this, I mean that if we

28

Chapter 4

have a set of events represented at di¨erent spatial locations, and impose an order on that set, then moving between the locations in that order generates a temporal sequence of the events. If speed of movement is constant and fast enough, then covering the space between event representations can reconstruct the time distribution of those events. Lashley (1951; see also chap. 2) argued that the events of animals' (including humans') experience are stored in the brain as a spatial distribution but are reproduced when required in a sequence from the spatial representation, thus creating the temporal distribution of the events, their serial order. A similar mechanism is thought to produce new sequences of behaviors (and thoughts) in a particular serial order that itself often has signi®cant meaning (as in speech). This (re)production of the temporal from the spatial may very well be a clue as to a useful way to think about the unfolding of behavior in time. 4.2

The Arrow of Time

Time is more than order and interval, however. As James (1890) and many others have pointed out, and everyone knows intuitively, there is a sense to time as experienced by humans (and perhaps other animals), a direction in which the order proceeds preferentially. We look out from the ``saddle point'' of the present, as James called it, remembering the past and predicting the future. We experience a sense of time passing, or ¯owing as a river ¯ows. Our very consciousness has been metaphorized as a ``stream'' ¯owing, and we all know that streams do not ¯ow uphill. Where does the directionality of time come from? The easy answer given by physics is the second law of thermodynamics: entropy always increases in an isolated system. Entropy, originally concerned with the ¯ow of heat between objects, is now interpreted as a measure of disorder, so that the second law now reads that isolated systems evolve toward a more disordered (more probable) state. For example, if we put a cold ice cube in a warm glass of cola, the ice cube will melt as the heat ¯ows from the cola and the glass to the ice, and the result will be a glass of cooler and somewhat diluted cola (but no ice). All of the ¯uid in the glass (and the glass itself, neglecting the surrounding air) will come to have the same temperature and the entropy of the system will be larger than at the moment when the ice cube entered the glass. In this explanation, the arrow of time represents the system's evolution from less probable (more organized) states to more probable (less organized) states. The passage of time is simply a statistical phenomenon. In a beautiful

Time

29

lecture on this question, however, Thomas Gold (1965) demonstrated that even these statistical tendencies do not su½ce. By simply reversing all processes in the universe, we can build a coherent picture in which time goes in the opposite direction; we are left with order without direction. Like Kant, Gold concluded that it is we who give time a direction through our experience of its ¯ow. He argued that this experience arises because memory is more accurate than prediction (simulation): we look out from the ``now'' in two directions that are asymmetrical, labeling the more accurate one to be what already happened, the less accurate to be what is yet to happen. This emphasis on memory, particularly ``episodic'' memory (Tulving 1983) is entirely consistent with James's ideas. What we remember feels more solid than what we think will happen in the future (but see Roediger 1996). The latter is usually generated by mental simulations (Kahneman and Tversky 1982), though occasionally we experience powerful and accurate ``intuitions'' about the future. Gold's explanation for time's arrow is consistent with another modern view of time as itself evolving along with the universe. Fraser (1982) argued that pure, symmetrical succession is the time of the physical world, whereas asymmetrical time (past, present, future) is the time of the biological and the psychological worlds. Just as the universe has evolved from the big bang singularity, so time has evolved from the atemporal (chaos of photons traveling at the speed of light), to the prototemporal (probabilistic speci®cation of temporal positions; no meaning of ``moment''; world of subatomic and atomic particles), to the eotemporal (simultaneities of chance; pure succession; ideal clock; directionless, symmetrical time; physicist's and dynamicist's t), to the biotemporal (simultaneities of need; physiological present; evolutionary or developmental past; predictive future; living organisms), and ®nally to the nootemporal (simultaneities of intent; mental present coexisting with physiological present; remembered past; imagined future). Of course all of these times exist now, in a nested hierarchical complex that is still evolving as humans (and perhaps other creatures) create ever more complex informational and interactional environments (cf. Hall 1983 and chap. 3). This notion is similar to one in which each dynamical system has its own characteristic cycle time for information exchange between elements (Alvarez de Lorenzana and Ward 1987), except that, for Fraser (1982), the very nature of time, not just its unit, is di¨erent at the di¨erent levels of the hierarchy. Fraser's view reminds us that eotemporal interval should not be taken to be the only property of time relevant to dynamical cognitive science. The directionality of biotemporal and nootemporal time can even help us

30

Chapter 4

unravel the temporal complexities of cognitive behavior (see ®g. 4.2 and chap. 5), and the existence of di¨erent time scales informs the frequency structure of temporally distributed ¯uctuations (chaps. 14±19). 4.3

Measuring Time

In spite of (or perhaps in reaction to) Feynman's statement in the second epigraph above, I have de®ned time in some detail, although, at least until dynamical cognitive science is more advanced, we must be content with a series of observations made at a series of moments of eotemporal clock time, on which we must impose somewhat arti®cially the directionality of nootemporal or biotemporal time. The clocks we use today are very accurate compared to the ®rst clocks used by Galileo. Nonetheless, those ®rst clocks were based on a fundamental principle that is important today. Galileo noticed that the period of a particular pendulum was always the same, even though the distance it traveled might change as the initial energy imparted to it was dissipated by friction (at least as measured by his pulse). A clock is merely a process that has a regular period, usually of movement between two states. Agreement among clocks is the standard by which we decide which are the best clocks. In cognitive science, we often use computer clocks based on oscillations in a crystal. Such oscillations are highly regular and can be very fast, making it possible to measure the timing of events to microseconds or even more precisely, though the signi®cant di¨erences cognitive scientists are interested in are usually on the order of 10 milliseconds (e.g., of reaction time). My rule of thumb is that measurement should be at least an order of magnitude more precise than the di¨erences I want to detect, so that time measurement error does not make it harder to detect the meaningful di¨erence, which usually means measuring time to the nearest millisecond. This should be su½ciently precise for most cognitive and behavioral time series, and is probably more precise than necessary for conscious cognitive behavior, which proceeds at about four operations per second (see chap. 16). Even if we measure time with a very precise clock, we often cannot observe behavior as precisely. A behavioral time series is composed of a series of observations that persists for a total time period, T, in which the observations are separated by an interval, Dt, both measured in, say, seconds of clock time. For example, consider recording, once per second, at which part of a complex painting I am looking. The period, T, puts a lower limit on the temporal regularities we can observe: if any occur more

Time

31

Figure 4.2 For eotemporal time either direction of events makes sense, whereas for biotemporal and nootemporal time only the top-to-bottom direction makes sense. (After Fraser 1982.)

32

Chapter 4

slowly than a frequency of 2/T we will not see them because a cycle begun during the observation period will not be completed while we are still collecting observations. For example, if we collected our looking observations for T ˆ 100 seconds, we would not notice a regular waxing and waning of my attention that caused me to gradually speed up my scanning behavior to a maximum, slow it down to a minimum, and so forth, repeating the cycle every 10 minutes (600 seconds). Such behavior would have a temporal frequency of 1=600 ˆ 0:0017 Hz. We would notice some change of scan rate over our observation period, but it would not be su½cient to tell us whether the change was regular or just a drift or trend for that one observation period (causing our time series to be nonstationary; see chaps. 11 and 12). Similarly, the sampling rate, the number of observations (samples) per second, or 1/Dt, imposes an upper limit on what we can observe in the time series: we cannot see regularities whose frequency is greater than 1=2Dt, or what is called the ``Nyquist frequency.'' If all relevant regularities occur below the Nyquist frequency, then we can be con®dent of having seen them all if only we sample for a long enough duration. If however, there are any regularities above that frequency, they will contaminate our time series in a process called ``aliasing,'' which arises from discretely sampling a continuous function at a rate that is too slow to capture all of its regular ¯uctuations. Although the e¨ects of aliasing can be dealt with or avoided (see Press et al. 1992), anyone dealing with time series should be aware of them and of the limitations of discrete sampling of a continuous process. Moreover, we should be mindful that quantum theory may sometimes apply even to cognitive behavior (see, for example, Hecht, Schlaer, and Pirenne 1942 on the absolute threshold of only six photons for vision). The quantum theory uncertainty principle for time, Dt ˆ h=DE, where E stands for energy and h is Planck's constant (6:6254  10ÿ27 erg sec), means we cannot simultaneously measure both time and energy for a physical system to any desired degree of accuracy. If we want more precision in knowing when something happened, we are forced to give up precision in measuring what it was that happened (energy change). Of course this minimum error is actually very small and would hinder neither classical physics nor ordinary cognitive science, where time is usually measured to the nearest millisecond. Nevertheless, it serves to remind us that, in measuring one aspect of a system, we may disturb that system, making measurement of another aspect less precise. If we think of time as generated by changes in large-scale systems the uncertainty principle might apply even to humans.

Time

33

Finally, there is the issue of continuity. Whereas, in physics and most dynamics, time is a continuous dimension, represented by the real numbers, discrete time might better characterize the unfolding of cognitive behavior in time, which could create discrete time instants, abrupt transitions between instants and places, like a Markov chain or a ®nite state machine transition between states with no intermediate positions. Although such an idea might seem inconsistent with the laws of physics, at least those of classical physics, it need not be. As Kilmister (1992, 58) puts it: ``All change, whether it be generation of a new element or incorporation of new information, takes place in discrete steps and the sequence of steps is the process.'' At least some physics can be reformulated in terms of the calculus of discrete di¨erences, in which the time derivative is the di¨erence between the values taken by a variable at two neighboring time instants, and di¨erentiation causes time shiftsÐa novel aspect that does not follow from di¨erentiation of a continuous variable (see, for example, Kau¨man and Noyes 1996). Finally, some of the most di½cult conundrums of modern physics, such as quantum gravity, can perhaps be understood in terms of a quantized or discretized time and space (t'Hooft 1989). How this view bears on dynamical cognitive science will be discussed in chapter 5.

This page intentionally left blank

Chapter 5 Cognitive Processes and Time

Time travels in divers paces with divers persons. ÐShakespeare, As You Like It We might . . . say this of time: It ambles with action, trots with thought, gallops with sensation. . . . ÐHerbert F. Crovitz, Galton's Walk ``How about me?'' he yelled, ``Am I to be just like a clockwork orange?'' ÐAnthony Burgess, A Clockwork Orange

5.1

Temporal Unfolding of Cognitive Behavior

In chapter 4, I set forth a constructive view of time, in which a cognitive process unfolds by progressing through a series of states, constructing time as it goes. In other words, the system changes. Rather than thinking of these changes as occurring in time, we can think of them as creating time itself. We do this by using the states as markers of time; we measure time by counting the number of state changes between recurrences of the same state. As Crovitz (1970) emphasized, following Galton, thoughts recur. Thought is cyclic. One way to construe this statement is by representing a cognitive system as a ®nite state machine. Tipler (1994) argued that humans can be modeled as ®nite state machines because they are quantum systems, and all such systems can be modeled as ®nite state machines. A ®nite state machine is simply a ®nite set of states and a transition matrix that records the probabilities of transitions between each state and each of the others. The Markov chains discussed in chapter 2 are examples of ®nite state machines. Clocks can be represented as ®nite state machines, usually with two states and with the probabilities of a transition from one state to the other near 1.0. A more complicated

36

Chapter 5

Figure 5.1 A simple ®nite state machine.

example is shown in ®gure 5.1, where the machine has four states (note that the transition probabilities are shown near the arrowheads), and the cyclic behavior is not at all clear. In particular, there is no absorbing state into which the system will collapse, and no guarantee that a certain state, once reached, will recur when the same preceding cycle of states recurs. And yet the system does cycle. There are only 4 4 ˆ 256 possible four-state cycles (including cycles like 3 ! 3 ! 3 ! 3), for example, and each must recur many times in a long sequence generated by the machine. The cyclicity is simply more exotic, the rhythm hard to discern from such a representation. Let us pretend that the system as diagrammed represents one spouse's cognitive and emotional state vis-a-vis the other as follows: 1P ˆ passion; 2I ˆ indi¨erence; 3A ˆ annoyance; 4M ˆ mild a¨ection. What does the diagram capture about the cyclic nature of this particular long-term relationship at a particular moment in time? It seems that, for the spouse under consideration, passion is a short-lived emotional state (probability of staying in that state is only 0.1, versus 0.9 for changing to another state), whereas indi¨erence (2I) is the most common state (sum of transition probabilities to that state is the greatest at 1.6). The short cycle of indi¨erence (2I) to annoyance (3A) and back again is quite common: the probabilities P…2I ! 3A† and P…3A ! 2I† are both high. Of course, the probabilities can themselves evolve. For example, Sternberg (1986) argued that passion is a common state at the beginning of a romantic relationship (all probabilities leading to passion are high, including the

Cognitive Processes and Time

37

Figure 5.2 Adaptation of the visual receptor complex. (From Ward 1991.)

probability to stay there, implying that probabilities leading away from passion are low), but the likelihood of passion wanes as the relationship progresses, eventually reaching a somewhat low state in most relationships (perhaps similar to the probabilities in ®gure 5.1). The ever-precessing sequence of cognitive and emotional states is only one example of the many aspects of cognitive behavior that demand description in terms of temporal unfolding. Although none is fully developed into a dynamical systems theory, let us brie¯y consider a few other compelling examples, beginning with information input. Everyone knows that sensory systems adapt, that the response of, say, the visual system to a patch of light striking the retina is greatest at ®rst and then rapidly wanes until a steady state is reached. If we think about this process as one of acquiring information, as Kenneth Norwich (1993) has, then adaptation represents a series of states of increasing knowledge about what is out there. Although we typically represent adaptation as continuous with time (see ®g. 5.2), according to Norwich's theory it is actually discrete, depending on a series of samplings of the stimulus energy by the receptor. The state of the receptor±ganglion cell complex before the light has struck the retina is one of ignorance about the light's intensity. After the light strikes, the receptor begins both to acquire information about its intensity and to lose information as it ``forgets'' the content of earlier

38

Chapter 5

samples. Eventually, a state is reached in which the forgetting exactly equals the acquiring: equilibrium. Any change in the light input disturbs this equilibrium, and causes the receptor to change its response. First the ®ring rate jumps to a new level above the previous equilibrium rate as the receptor±ganglion cell complex enters the state of ignorance of the current light intensity, and then it falls, through a series of intermediate states representing information acquisition, to a new equilibrium state that represents the balance of information acquisition and loss for the new light intensity. Adaptation is just one of the many processes involved in the microgenesis of a visual percept (see, for example, Kalmus and Bachmann 1980), which also includes the time course of feature extraction, object formation, categorization, identi®cation, comparison with other perceptual objects, application of semantic rules, evocation of associations, and so forth. These processes have been studied repeatedly and thoroughly in every perceptual system, but seldom with the aim of completely describing the temporal unfolding of the percept. One particular time-critical aspect of visual perception is the sequence of eye movements made when scanning a scene. Figure 5.3 shows how these movements occur even when scanning a scene at rest; they are much more critical when scanning a scene in motion, as when driving a vehicle or viewing a ®lm. Enormous e¨orts have been expended in understanding eye movements, although usually within certain contexts, such as reading (e.g., Just and Carpenter 1976, 1987), driving, ¯ying, or searching a scene, with an emphasis on the information extracted rather than on the complex interplay of that information with the process of scanning itself. If the eye movement system were to be represented as a ®nite state machine, with each state de®ned by the part of a given scene focused on, then it is clear that di¨erent scan patterns could be described by di¨erent state transition matrices, and that such matrices would both depend on the preconditions of a particular situation and dynamically (and complexly) on the information being extracted. Eye movements in a situation in which the scene itself is also changing, perhaps as a result of movement of the body in response to the information extracted from the previous scene, would be described by a sequence of transition matrices generated by the interaction of the extracted information and the goals of the person. Many more examples of temporal unfolding in cognition could be adduced. Learning is obviously a time-critical process, including both timing aspects (see sec. 5.2) and unfolding as the knowledge or skill state of the organism changes. Memory, the record of that learning, is also

Cognitive Processes and Time

Figure 5.3 Eye movements (bottom) while scanning a picture (top). (After Yarbus 1967.)

39

40

Chapter 5

dynamical (see sec. 5.3). But beyond these obvious examples are others. Indeed thinking itself cycles, creates time, and unfolds in the time it creates. The exercise of logic is sequence critical, and thus time critical. A mathematical or logical proof only works if the steps are presented and cognized in the correct sequence. Although a proof can be modeled by a degenerate ®nite state machine, what is more interesting is a human's understanding of a proof, which can be modeled by a much more complicated ®nite state machine, with many intermediary states and probably changing matrices of transition probabilities (memory). Problem solving also has its own rhythm, which can be arrived at by asking people to verbalize their thinking during problem solving (e.g., Ericsson and Simon 1993), or by having them make judgments about how close they might be to a solution. Metcalfe 1986 and Metcalfe and Wiebe 1987 on subjective ``feeling of knowing'' are two such studies. In Metcalfe 1986, subjects solving insight problems were actually less likely to give the correct solution when they felt ``hot'' (i.e., the solution was imminent) than when they felt only ``lukewarm'' (not very close to a solution but on the right track). Metcalfe and Wiebe 1987 showed that this pattern holds only for insight problems; for algebra problems, subjects' feeling that they were very close to a solution did indeed correctly predict their subsequently obtaining that solution. In both studies, time series of judgments of feelings of ``warmth'' were collected as subjects solved problems. Although the focus of the studies was on the inherently temporal aspect of predicting performance (both by the subjects themselves and by normative ratings), the analysis was not extended to modeling the full time series. It would be informative to make a ®nite state machine model of those time series, in which the probability of switching from the various states of warmth into other such states and into the absorbing ``problem solved'' state could be estimated from the transitions occurring in the various empirical time series. Such a model, by displaying the overall dynamical properties of the process of solving insight problems, could perhaps suggest additional hypotheses about the mechanisms involved. 5.2

Timing of Cognitive Behavior

Many cognitive science phenomena seem to involve a di¨erent aspect of time, what we might call ``timing,'' where the occurrence of stimuli or states at particular times and places seems to matter the most. For example, in visual masking (e.g., Kahneman 1968; Enns and Dilollo 2000), given appropriate spatial relations, it is exactly when the masking stimulus

Cognitive Processes and Time

41

occurs with respect to the onset and/or o¨set of the target stimulus that determines the masking e¨ect. If the two overlap in time, then brightness masking occurs. If the target precedes the mask by about 60±100 msec, then maximal backward masking occurs; if vice versa, then forward masking. And if target and mask are more than 200 msec apart in time, then usually no masking occurs at all. The very names of the phenomena are informed by the temporal relationship between the stimuli. Similar examples occur in the orienting of attention (e.g., Wright and Ward 1998), inhibition of return (e.g., Klein 2000), priming (e.g., Houghton and Tipper 1994), alerting (e.g., Sokolov 1975), and warning (e.g., Woodworth 1938), among many other processes. But timing actually involves synchronizing two or more processes unfolding in time. In masking, for example, the microgenesis of the percept of the target stimulus interacts with that of the masking stimulus at a critical moment in that of the target. If the target percept is already fully formed, it is perceived. When masking occurs, the percept of the mask has undermined the processes leading to the perception of the target. The situation is similar in the other timing examples: timing is always about the relationship between two or more unfoldings in time. 5.3

Memory

Memory would seem to be the cognitive process that most directly addresses time. Memory di¨erentiates the past from the present, and the di¨erences between the ``feeling'' of a memory and the feeling of an anticipation might help to di¨erentiate both from the future (see chap. 4). In discussing memory, we suppose that the cognitive system is always changing and that current inputs have lasting e¨ects that may wax (learning, consolidating) or wane (relaxing, forgetting) over time. What is the nature of this inherently temporal process? One axiom of the modern understanding of memory is that it is constructive, and therefore inherently dynamical. Memories are not simply records sitting on a library shelf, waiting for a librarian to take them o¨ and bring them out into the bright light of working memory. When we remember an episode from our past, we actively build the ``memory'' from traces of perceptual and cognitive activations of the brain that were triggered by the episode we are remembering (literally ``re-bringing to mind''). These old traces are recycled in the current context, with current activations contributing strongly to the remembered experience. Although it is useful to classify memory into various types, including sensory, pro-

42

Chapter 5

cedural, episodic, gist, implicit, explicit, working, and repisodic, each type abides by the same underlying principle: changes in the brain at one time a¨ect the activity of the brain at a later time (see Squire 1987). Indeed, this principle can be extended to all forms of ``memory,'' including computer memory, genetic memory, and even more abstract systemic memory, implying that there might be a universal model for memory that would illuminate the human case (Clancey 1999; Minelli 1971), possibly in the form of a ®nite state machine, where memory processes would be modeled by changing of state transition probabilities. In the absence of such a universal model, we must work at modeling the memory states of the human or other animal brain. Because the activation sequence of the entire brain is too complex to model as a ®nite state machine at the neural level or lower, we will have to be content with models at a fairly high level for the present. A few such models are already available to serve as a foundation, for example, network models of associative memory (e.g., Anderson 1976; Rinkus 1996). Many models of working memory have been proposed, although not at the same level of detail. A particularly intriguing one conceptualizes working memory as a push-down stack, much as some memory registers are organized in computer operating systems. Similar models could be created for processes of cognitive development, both for the learning of thinking skills (including memory skills) and for the learning of facts. In these models, ®nite memory states should not be confused with the ``stages'' of the various stage theories of cognitive development, such as Piaget's (1954). One particularly interesting cognitive ability to model would be memory for sequence. As I discuss in chapters 28±30, our perception of sequence is biased, and at least part of this bias depends on our memory for the order in which recent events have happened. There is a considerable, though sporadic, literature on sequential aspects of memory. One interesting set of recent studies (Houston 1999; Wegner, Quillian, and Houston 1996) di¨erentiated the e¨ect of instructions to suppress memory for a ®lm, on memory of the sequence of events from that on memory for the events themselves. Another recent study (Woltz et al. 1996) showed that general sequence memory was more useful than speci®c sequence memory in skill acquisition. A third (Clancey 1999) describes in detail an inherently dynamical theory of memory based on the distinction between simultaneous and sequential activations of concepts in the brain (see chap. 3). In general, memory (along with most other cognitive phenomena) exists at many timescales, from very short (1 msec or less; physical and chemical processes in the brain), to fairly short (25 msec; memory scanning),

Cognitive Processes and Time

43

to medium (250 msec; conscious operations), to fairly long (seconds, minutes; short-term or working memory), to very long (hours, days, years; episodic memory) to longer still, (centuries; scienti®c research, organizational memory), to evolutionary (millennia; evolution of cognitive processes). In particular, more precise modeling of explicit, autobiographical memory systems, including their evolution, could contribute greatly to the solution of the problem of consciousness (see Donald 1995). Although some of these topics are taken up in later chapters, many others must await the future e¨orts of dynamical modelers.

This page intentionally left blank

Chapter 6 Systems and General Systems Theory

What is a system? As any poet knows, a system is a way of looking at the world. ÐGerald M. Weinberg, An Introduction to General Systems Thinking There exist . . . systems [whose] quantization in space encompasses their structure; their quantization in time represents their orbits. ÐArthur S. Iberall, Toward a General Science of Viable Systems

Some while after the Beginning, there was Newtonian science. And Ludwig von Bertalan¨y said, ``Let the whole system be greater than the sum of its parts.'' Norbert Wiener added, ``Let negative and positive feedback ¯ow throughout the system.'' Ross Ashby whispered, ``Let the system have the requisite amount of variety to manage its environment.'' And general systems theory was born. This book could be viewed as an attempt to embed cognitive science ®rmly in a general, dynamical, systems framework. Many of the concepts used thus far, and of those used later on, can be traced to fundamental general systems theory (for a ®ne introduction, with articles by many of the most important practitioners, see Klir 1972a). General systems theory embodies many of the formal theories whose use I advocate in analyzing cognitive behavior, including the theory of ®nite state and stochastic machines, Turing machines, and the like. Ironically, however, systems theory remains a somewhat obscure discipline, usually taught in engineering schools and occasionally also in business schools. Some of its concepts have been co-opted by the more modern and fashionable ``complexity theory'' (e.g., Waldrop 1992). As we will see, systems concepts are crucial to the development of dynamical cognitive science. I can only introduce a few of the major ideas here and in the following chapters, leaving readers who wish to apply these ideas in their respective contexts to pursue them in greater detail elsewhere.

46

6.1

Chapter 6

Systems

What is a system and why is it relevant to dynamical cognitive science? As Weinberg (1975, 52) says in his excellent introduction to general systems theory, and in the ®rst epigraph above, ``a system is a way of looking at the world.'' A point of view is necessary because even that tiny part of the world, the human brain, is complex, containing around 100 billion neurons and probably more than 10 trillion synapses. Indeed, even when we try to limit our attention to a single cognitive system, we run into complexity that demands a point of view. One way of coming to grips with this problem is to contemplate the ``Bremermann limit'' (Bremermann 1962). At approximately 10 93 bits (a bit is a Binary digIT, 0 or 1, equivalent to a Boolean logic truth value, a unit of Shannon information; see Shannon and Weaver 1949), the Bremermann limit represents the processing capacity of a digital computer the size of the Earth (6  10 27 grams) over the time the Earth has existed (10 10 years at about 3:14  10 7 sec/ year ˆ 3:14  10 17 sec), assuming a maximum processing rate of 2  10 47 bits/sec/gram. Many reasonable computation problems are beyond the Bremermann limit. For example, even a simple problem in Boolean algebra (the logic implemented by general-purpose computers and universal Turing machines) can quickly exceed it. Klir (1972b) gives an example of implementing on a computer a function that maps a set of n logic (0 or 1) input variables to a set of n logic output variables using what are called ``universal elements'' or computing modules. He shows that the number of bits involved in determining how to implement this function for various n‡1 numbers of input variables, n, is N ˆ n
2 2 . For n ˆ 1, N ˆ 16, but 155 for n ˆ 8, N ˆ 10 and for n ˆ 10, N ˆ 10 617 Ðthe latter is more than 500 orders of magnitude greater than the Bremermann limit. (Special theoretical methods are used to solve these and related problems in computer design.) Because similar computational problems arise in science, it is clear we must simplify. And whenever we simplify, we implement a point of view. Building a model or a scienti®c theory (see chap. 7) is a way of simplifying, of making the otherwise overwhelming complexity manageable. Thus a system is a theoretical construct that simpli®es nature. Nonetheless, the system must be complex enough to manage relevant aspects of its environment. For example, in a human memory system, there must be enough complexity (or what Ashby called ``variety'') to deal with remembering the things a human has to remember (grocery lists, words of a language, faces of relatives and friends, rules of games). Ashby (1958)

Systems and General Systems Theory

47

called this principle the ``law of requisite variety.'' It was extended by many, including Beer (1979), who pointed out that there are various ways a system could amplify its own variety to cope with the much larger amount of relevant variety from the environment. Finally, Alvarez de Lorenzana and Ward (1987) argued that systems should be placed into an evolutionary context, in which those with ever-increasing variety emerge from an environment also increasing in complexity. They identi®ed two principles by which a system evolves: combinatorial expansion, a linear process whereby the system develops the unrealized potential in its fundamental properties; and generative condensation, a nonlinear process whereby complexes of the elements of the old system recombine to create fundamental elements of a new system with new possibilities. Both processes operate through system-environment interaction. They argued that, through successive stages of combinatorial expansion and generative condensation, the universe has evolved from the ®rst, simple physical systems (such as the hydrogen atom), through ever-increasing complexity of physical, chemical, biochemical, biological, and ®nally social and intellectual systems into the world we experience today. Each of these system levels can be described by its own fundamental properties, including units of time and space unique to the interactions taking place between the system elements at that level (see chap. 4). In this way, a cognitive system is no di¨erent from others, and should be characterized by the relevant fundamental systemic properties, such as time and space units, information input capacity, and potential for combinatorial expansion. One principle that makes some systems easier to deal with is the ``square root of n law'': variances of statistics obtained by sampling from probability distributions decrease as the sample size increases (see SchroÈdinger 1944/1967). For example, the standard deviation of the distribution of means of random samples taken from any probability distribution (the standard error of the mean) is equal to the population standard deviap tion divided by n. For very large systems, then, we can ®nd statistical regularities. On the other hand, for very small systems, such as simple machines, we can successfully analyze behavior in terms of the interactions of their individual components. For systems of medium size, however, neither approach works well; in such systems, we observe ¯uctuations of many sizes, irregularities and lack of predictability, and great theoretical di½culty. Unfortunately, systems of medium size, such as cognitive systems and even brains (for some purposes at least), are the rule. With medium-size systems, we seldom bene®t from the square root of n law: the simpli®cation we must undertake will cause us to omit many important

48

Chapter 6

Figure 6.1 Where general systems thinking ®ts in. (From Weinberg 1975, p. 18. ( 2001, 1975 by Gerald M. Weinberg. All rights reserved. Reprinted by permission of Dorset House Publishing.)

elements, relations, or both. Thus we can expect deviations from theoretical predictions to occur with regularity, whereas regularity in system behavior will seldom be seen. Figure 6.1 illustrates what we must deal with in most medium-size systems, and especially in cognitive systemsÐ organized complexity, the most di½cult type of system to understand, but also potentially the most rewarding. 6.2

General Systems Theory

It is not enough to acknowledge that we are dealing with systems of medium size and then go on to do our usual science. As we simplify and search for statistical and other regularities, we must be constantly aware that we have simpli®ed, and that there are other points of view not represented by the one we are taking. Other researchers will take some of those points of view, and what might emerge eventually is a kind of order of the second degree: laws about the laws of what we are studying. Those who have applied general systems theory across the various ®elds of science and engineering have come up with quite a few laws about laws. Some are useful and others appear whimsical, but all contain a grain of truth about how to deal with organized complexity, the realm in which we

Systems and General Systems Theory

49

Table 6.1 General Systems Laws (Laws about Laws) Name

Content

Composition law (Weinberg)

The whole is more than the sum of its parts.

Decomposition law (Weinberg)

The part is more than a fraction of the whole.

Principle of indi¨erence (Weinberg)

Laws should not depend on a particular choice of symbols.

Principle of di¨erence (Weinberg)

Laws should not depend on a particular choice of symbols, but they usually do.

Principle of invariance (Weinberg)

Change can be understood only by observing what remains invariant, and permanence only by what is transformed.

Fundamental theorem (Gall)

New systems mean new problems.

Law of large systems (Gall)

Large systems usually operate in failure mode.

Law of function (Gall)

Systems tend to oppose their own proper function.

Law of growth (Gall)

Systems grow.

Sources: Weinberg 1975 and Gall 1975.

nearly always ®nd ourselves in cognitive science. Table 6.1 lists a few such general system laws, just to reveal their ¯avor. Probably the most famous of the general systems laws listed in table 6.1 is the ``composition law'': the whole is more than the sum of its parts. This law refers to what has been called ``emergence,'' the idea that a simple combination of the properties of the separate elements of a system does not describe the properties of the system itself. One famous example is that of the water molecule, H2 O. Water molecules freeze at 0  C, whereas both oxygen and hydrogen, its constituents, are gases at that temperature. Water does not become a gas until 100  C. We now understand why water has these properties (and others that do not match those of hydrogen and oxygen): they arise because of the unique geometry of electrical charge created by the hydrogen and oxygen atoms when joined into a water molecule. This geometry is now understood in terms of quantum electrodynamics, but it is not a simple linear combination of properties of hydrogen and oxygen atoms. It is the relationship between its constituent atoms that create water's properties; its atoms must be

50

Chapter 6

joined in just the right way. Thus water's properties are said to be ``emergent'' because they cannot be obtained directly from the properties of its constituents. Cognitive systems often seem to exhibit ``emergence''; indeed, time and again, the most elusive cognitive phenomenon of all, human consciousness, has been said to be ``emergent'' relative to the human brain. A ®nal important distinction in this brief introduction to general systems theory is that between open and closed systems. The ideal system of thermodynamics and science in general is closed, meaning that all in¯uences on system behavior are known and either under control or fully described. In a closed system, entropy (or disorder) always increases with time. In contrast, all real systems are open, meaning that they are in contact with a larger environment that produces uncontrolled and possibly undescribed inputs. The Earth, which receives some of the extravagant outpouring of energy from the Sun, is just such an open system, as are the many constituent systems on its surface. Some systems, such as those in physics experiments using particle accelerators, can be su½ciently isolated from the larger environment that fairly simple descriptions of their behavior will be accurate. Open system are the rule, however, and their behavior re¯ects this openness. The most interesting open systems are those far from equilibrium, in which some of the energy input is used to maintain self-organized patterns: thermal convection cells, for example, or exotic chemical reactions such as the Belusov-Zhabotinski reaction, or possibly even human consciousness (see Prigogine and Stengers 1984; Kelso 1995). 6.3

Dynamical Systems Theory

Dynamical systems theory attempts to describe the temporal unfolding of a system. It is concerned with two fundamental concepts, those of change and time, both of which I have discussed in earlier chapters. There are two standard ways to mathematically describe a dynamical system (von Bertelan¨y 1972): an external description, which consists of a set of transfer functions, usually expressed as systems of linear equations relating inputs to outputs; and an internal description, which consists of a set of n simultaneous partial di¨erential ``equations of motion,'' dYi ˆ fi …Y1 ; Y2 ; . . . Yn †; dt

…6:1†

where Yi represents the system state variables and i ˆ 1; 2; . . . n. (If your

Systems and General Systems Theory

51

mathematical training did not include calculus, be assured there will be gentler and more helpful introductions to such equations in chapter 8.) The important thing to note here is the ``dt'' in the denominator of the ®rst term, which means that the equations are expressing the rate of change of the system state variables, Y1 ; Y2 ; . . . Yn , with respect to time, t. Although this is standard physical (eotemporal) time, depending on the state variables and the time measurement, these ``equations of motion'' can become very exotic indeed. However di½cult they may be to work with, we need to face them, for they formally describe important system properties, such as wholeness and sum, growth, competition, and equilibrium (von Bertelan¨y 1968). In doing so, they describe the ``motion'' of the system in state space, an n-dimensional space in which the current state of the system is represented as a vector (one value for each state variable) and in which the changes of values of these state variables (i.e., the changes in the system's state) are represented as trajectories, or paths. Trajectories can be divided into categories, such as asymptotically stable (all trajectories su½ciently close to a particular one approach it asymptotically with time), neutrally stable (all trajectories su½ciently close to a particular one stay in its vicinity), and unstable (all trajectories su½ciently close to a particular one diverge from it with time). When the equations of motion de®ned by equation 6.1 are linear, many things can be said about the system described by them. Systems described by nonlinear di¨erential equations, though they often exhibit dramatic instabilities and are harder to understand, are nevertheless the most interesting systems, and the ones on which we will concentrate in this book. After a brief diversion to establish some important concepts about theories, models, and scienti®c understanding in chapter 7, we will return to the development of a dynamical system theory approach to cognitive science in chapter 8.

This page intentionally left blank

Chapter 7 Science and Theory

Even the artists appreciate sunsets, and the ocean waves, and the march of the stars across the heavens. . . . As we look into these things we get an aesthetic pleasure from them directly on observation. There is also a rhythm and a pattern between the phenomena of nature which is not apparent to the eye, but only to the eye of analysis; and it is these rhythms and patterns which we call Physical Laws. ÐRichard Phillips Feynman, QED

Why should I presume to tell you, my scientist readers, what I think science is? There are two reasons. First, I simply cannot resist the natural temptation to inject a few memes (Dawkins 1976) on this topic into your brains. Second, you need to understand what I am trying to do, even if you do not agree with me. Knowing the framework within which I do science, you can compare it with your own and judge accordingly. I should tell you that I have been heavily in¯uenced by physics, especially by quantum mechanics, through a few of its practitioners most willing to put themselves on public display, such as Richard Feynman (1965, 1985) and Erwin SchroÈdinger (1944/1967). 7.1

The Mandala of Science

Michael Ovenden was a very good astrophysicist, colleague, and friend of mine, who died too young at 62. We often argued about what science is and how it should be used, before, during, and after the meetings of the General and Applied Systems Workshop at the University of British Columbia, which we both regularly attended and to which we both regularly contributed. In Michael's highly rigorous view, only physics (perhaps including astrophysics) had achieved the status of a true science. Figure 7.1 shows my version of the basic elements of his view. (In Michael's version, there was only one arrow, labeled ``intuition,'' between ``theory''

54

Chapter 7

Figure 7.1 Ovenden-Ward mandala of science.

and ``empirical generalizations,'' and it went in the direction of the arrow I have labeled ``creative thought.'') Michael held a dualist stance on the relation between mind and matter, believing that the inspiration to create a useful scienti®c theory came from a mysterious place to which we had no access, not even in principle. Thus he rejected the idea that psychology would or could ever become a science. Although Michael never convinced me to eschew psychology, he did in¯uence the way I think about science, along the lines displayed in ®gure 7.1, whose mandala divides science into products (theories, empirical generalizations, and empirical observations) and processes (deduction, induction, and creative thought). Cognitive scientists study the processes of science; we and other scientists can of course use them without understanding their mechanisms. These ideals of deduction and induction (arguing from the general to the speci®c and vice versa) are described precisely and applied rigorously by mathematicians and logicians in the context of proving theorems in formal systems. As practiced by scientists, however, they can range from loose arguments and wild speculations to

Science and Theory

55

precise mathematical or other formal derivations. Creative thought is still somewhat of a mystery, although we have studied it for many years. Some (e.g., De Bono 1970) think it requires some nonlinearity and others (e.g., Crovitz 1970) argue for a random element. The ability to make analogies and sophisticated guesses, to see patterns and relations, to hold many facts and concepts simultaneously before the mind's eye, and unusual perseveranceÐall seem to contribute to successful creative thought. The mandala emphasizes that science is a system, a whole that is different from the sum of its parts. Indeed, the necessity of all parts and the su½ciency of none are why Ovenden thought that only physics was a true science. He argued that biology, chemistry (though he wavered here), psychology, and other ``sciences'' were not true sciences chie¯y because they lacked deep formal theories whose precisely veri®able quantitative predictions had been extensively tested. Physics does possess such theories, for example, quantum electrodynamics (see Feynman 1985), quantum mechanics as a whole (see Baggott 1992), and general relativity (Einstein 1961; Thorne 1994). In biology, Darwin's theory of evolution by natural selection, now quite modi®ed and sophisticated, ®nds itself in nearly every modern biologist's toolbox, yet profound theoretical problems remain (see, for example, Kau¨man 1993). The remarkable ``central dogma'' of molecular biology, that DNA and its associated molecules both carry and implement the genetic code that informs the variety of life on this planet, is fairly well accepted, yet there is still no comprehensive theory about how the information encoded in DNA interacts with its environment to form and regulate a life-form (see Nass 1970). Cognitive science abounds with theories of perception, memory, reasoning, language production, and understanding, many of which are comprehensive and quantitative, such as the entropy theory of perception (Norwich 1993) and the wave theory of psychophysics (Link 1992), or comprehensive and formal in a di¨erent way, such as Newell's SOAR approach (1990) to cognition. Yet none has attained anything near the scienti®c rigor of quantum mechanics or general relativity. If we are to more closely approach the impressive accomplishments of physics, we must incorporate a dynamical component into cognitive theory, and into psychological theory as a whole. And because of the complex systemic nature of cognitive and psychological phenomena, the dynamical component will have to be in the form of dynamical systems theory. This book aims to convince at least some of its readers on both these points.

56

7.2

Chapter 7

Formal Theories

Although theories come in many types, I favor formal theories, a bias shared by at least some others in cognitive science. A formal theory is distinguished by its reliance on a few, precisely stated, fundamental principles from which rigorous application of logical manipulations can then produce a large variety of useful implications. Maxwell's theory of electromagnetism is an elegant and important, though di½cult to understand, example of a formal theory in physics. Basically, what Maxwell did was to transform Faraday's ideas about electrical and magnetic lines of force into a precise mathematical theory (see Born 1962; Wise 1979). The theory's fundamental axioms are written in the form of four partial di¨erential equations: `
E ˆ r=e

…7:1†

`
Bˆ0

…7:2†

`Eˆÿ

qB qt

`  B ˆ msE ‡ me

…7:3† qE qt

…7:4†

In these equations, E stands for the electrical ®eld intensity, B for the magnetic ®eld intensity, r for electrical charge density, e for the dielectric constant, m for the atomic weight (relative to that of hydrogen) of the electrical conductor (e.g., a copper wire), and s for the conductivity of the electrical conductor (values for copper, silver, gold, and platinum are high). Finally, q=qt, the partial di¨erential operator, means to take the partial derivative of the relevant multivariable function in the numerator (in this case, E or B) with respect to the variable in the denominator (in this case, t, or eotemporal time); here, we are interested in the rate of change of E or B with respect to time. E and B are vector-valued functions of space, that is, each ``value'' of the function is a vector in geometric space (speci®ed by coordinates x; y; z). Vectors have both a direction in space (the orientation of the vector with respect to the axes of the space) and a length, or magnitude (how far from the originÐ0,0,0Ð the end of the vector is). The symbol ` related to E and B by the raised dot in the ®rst two equations stands for the di¨erential operator ``divergence'' or simply ``div,'' the scalar product of the operator (a partial differential expression in the spatial coordinate vectors) with the relevant vector-valued function, in this case, E or B. Scalar products of vectors

Science and Theory

57

produce a scalar, or nonvector, number, at each point in space. The symbol ` related to E and B by the ``D'' in the other two equations stands for the di¨erential operator mathematicians call ``curl,'' the vector product between the operator and the relevant vector-valued function, which produces another vector at each point in space (instead of a scalar, as with ``div''). It is not necessary to understand these equations completely to grasp their beauty and fundamental importance. Equations 7.1 and 7.2 describe how electricity and magnetism behave separately. In words, equation 7.1 means that whenever an electrical charge (r=e) occurs, an electrical ®eld (E) appears that exactly balances the charge everywhere in space. Equation 7.2 means that a magnetic ®eld (B) is not accompanied by any free magnetic ``charges''; such charges do not exist. Equations 7.3 and 7.4, which contain the curl operator, describe how electricity and magnetism interact. Equation 7.3 states that every magnetic ``current'' created by the displacement or disparity of magnetic force lines in space induces an electric ®eld in a perpendicular direction (minus sign on the right). Similarly, equation 7.4 states that every electrical current, including one made up entirely of the displacement or disparity of electrical charge density in a vacuum, induces a magnetic ®eld. When the values of the constants representative of the vacuum are entered into equations 7.3 and 7.4, the equations describe the mutually generating propagation of the electrical and magnetic ®elds that make up a ``light wave'' or photon. Notice especially that the theory consists of a set of partial di¨erential equations in time and space variables. The equations describe the distribution of electrical and magnetic force over space and how these distributions change in time. This (relatively) simple theory thus describes all electrical, magnetic, and electromagnetic phenomena, including those of light itself, with incredible precision. One of the crowning achievements of Einstein's special theory of relativity was that Maxwell's equations could be derived in it. If they had not been, Einstein's theory would have been incorrect. In physics, the formal, mathematical statement is the theory, and any accompanying verbal or pictorial explication is secondary, which makes physics di½cult even for the initiated because mathematics requires careful and often deep thoughtÐand virtually impossible for the uninitiated because the combination of jargon and di½cult mathematics is impenetrable. This situation is unfortunate, but in my opinion the advantages of formal theory outweigh its disadvantages. These advantages are spelled out quite nicely by Feynman (1965), and include both the enormous simplifying and precision-inducing character of mathematical expression, and

58

Chapter 7

the bringing to bear of the entire reasoning framework of mathematics on the problem at hand. As Feynman wrote, ``it is impossible to explain honestly the beauties of the laws of nature in a way that people can feel, without their having some deep understanding of mathematics. I am sorry, but this seems to be the case'' (pp. 39±40). But cognitive science is not physics, and I know many will object that, in our context, formal theory is too restrictive and the simpli®cations required too radical, that informal theories are richer, admit a wider variety of interpretations capturing a wider range of phenomena, and so forth. Many of these objections are well taken. As Platt (1964, 352) put it: ``The logical net is coarse but strong, the mathematical net is ®ne but weak.'' By the ``logical net'' he meant an informal but correct understanding of a phenomenon, expressed in words or perhaps in ¯ow or structural diagrams, with perhaps an equation or three thrown in. Such an understanding is necessary before proceeding to a formal statement: although it is ``coarse'' and cannot make detailed quantitative predictions, it is nevertheless ``strong,'' in that it does indeed capture the essence of the understanding. In this sense, other scientists, especially Faraday, had understood the relation between electricity and magnetism somewhat before Maxwell wrote his equations, and had even expressed some of their understanding formally, as equations summarizing empirical generalizations (e.g., Ohm's law, E ˆ IR). Thus, too, the mathematical theory produced by Maxwell when he wrote his four simple equations was ``®ne'' because it made precise quantitative predictions about a vast range of electromagnetic phenomena, including those of light waves, but it was ``weak,'' in that if any of those predictions proved wrong, to even the slightest extent, the theory was discon®rmed. Many people will nevertheless feel that Faraday's pictures of interacting electrical and magnetic ®elds, though they do not do the work that Maxwell's equations do, give them a better grasp of electromagnetism than do Maxwell's equations. It is possible to be formal without being mathematical. Gregg and Simon (1967) argued that some sort of symbolic information-processing language might be more appropriate than mathematics for cognitive science. They showed how Trabasso and Bower (1970) in formulating their Markov chain theory of concept learning, often had to resort to informal theory because mathematical theory failed to apply to some important, empirical phenomenon; how, when they recast the informal theory into a formal symbolic language, much like a computer programming language, it captured the concept learning phenomena quite well

Science and Theory

59

while avoiding the inconsistency and imprecision usually associated with informal theory; and how they could mathematize the formal informationprocessing theory whenever necessary to generate and test quantitative predictions. Such an approach is certainly still a viable one to pursue in cognitive science; an excellent example is Newell's SOAR theory (1990). Although quantitative precision has always been and still is the goal of the physical sciences, when we move into the realm of nonlinear dynamics and cognitive science, such precision may not be generally attainable. Systems described by nonlinear di¨erential equations often have an extreme sensitivity to initial conditions: even a small amount of error in the speci®cation of the initial or current conditions of such a system will cause the predicted trajectory of the system to diverge radically from the actual trajectory. Although the system might end up in the predicted region of the system state space, its precise state at any given time will not be predictable with the precision required by most physical theories. This situation is called ``chaos'' (see chap. 23). Discussing the consequences of this lack of predictability for sciences that deal with nonlinear dynamics, Kellert (1993) argued that we might need to adopt a new approach to these sciences, in which the prediction of patterns of change would replace the prediction of precise values of system state variables. In such an approach, formal theories would be just as desirable because the need for consistency and precision of ideas would be just as great, but the goals of these theories would be accurate predictions of patterns of change, rather than of change magnitudes. Because system dynamics are often best studied as geometrical ®gures, the emphasis may well become more geometric. Nevertheless, the goal of precise and deep understanding remains intact. 7.3

Principle of Complementarity

According to Weinberg's principle of complementarity (1975, 120): ``Any two points of view are complementary.'' No one point of view provides the entire understanding. Sometimes di¨erent points of view are contradictoryÐthat is just the nature of the world. Paradoxes will arise and may be resolved, but there will always be multiple points of view. Two good examples are the wave/particle duality of light (photons; see chap. 13) and the Heisenberg uncertainty principle (measurements of one aspect of a system interfere with knowledge of other aspects). In the context of dynamical systems theory, Weinberg's principle reminds us that, having

60

Chapter 7

had to simplify, to adopt a point of view, and to accept some uncertainty about a system's behavior, we can take other points of view that, provided they are of equal rigor and depth, will undoubtedly contribute different elements to our understanding of the system we are studying. This is not a license to abandon scienti®c rigor and speculate wildly, but simply a reminder that sometimes, as in the wave/particle duality of light, complementary points of view can all be useful, even though each is incomplete. I expect this will often be the case in dynamical cognitive science.

Chapter 8 Dynamical versus Statical Models

[Re] Zeno's argument of the arrow[:] The natural view . . . is that there are things which change [but] philosophers have developed two paradoxes. . . . How ridiculous to say there is no arrow! say the ``static'' party. How ridiculous to say there is no ¯ight! say the ``dynamic'' party. The unfortunate man who stands in the middle and maintains that there is both the arrow and its ¯ight is assumed by the disputants to deny both; he is therefore pierced, like Saint Sebastian, by the arrow from one side and by its ¯ight from the other. ÐBertrand Russell, A History of Western Civilization Nothing is permanent except change. ÐHeraclitus

8.1

Theories, Models, and Data

Theories are grand things, even informal theories. Ideally, they are general and yet parsimonious, broadly applicable and yet succinct. A theory is a system within which to think about many speci®c phenomena. Again, we can take our pick from several prominent examples in physics, such as general relativity or quantum mechanics. Within these theories, models of more speci®c situations can be created, for example, a model of how light interacts with a sheet of glass in quantum mechanics (actually, quantum electrodynamics; see Feynman 1985). Psychology does have a few such theories, a prominent example being signal detection theory. Itself an application of statistical decision theory to detection of signals by radio receivers, this engineering theory was applied to human performance and has since become very widely used as an approach to measuring performance independent of motivational biases (e.g., Green and Swets 1966; Macmillan and Creelman 1991). Many models have been created using the theoretical constructs of signal detection theory. A prominent example is the excellent model of auditory intensity perception developed by

62

Chapter 8

Braida and Durlach (1988). In both physics and psychology, speci®c models are consistent with a particular, more general theory and use its terminology, but are actually applications of the theory in a restricted domain. Of course, models also can be created in the absence of theory. In this case, although it holds true that the elements and relations in the model system (often a mathematical system) represent the elements and relations of the actual system being modeled (often an empirical system), no generally accepted theoretical principles, or only broad, informal theoretical frameworks, are invoked to connect the model with a wider domain of phenomena. This is common practice in psychology, especially in domains like perception and cognition, where much progress was made from the 1960s through the 1980s with information-processing models of restricted domains such as attention, speech, judgment, and decision making. For this reason, I will be referring mostly to models, although my (perhaps distant) goal remains the development of a general theoretical system within which cognitive phenomena can be modeled. Some candidate systems have been proposed, for example, Newell's SOAR system (1990), Anderson, Matessa, and Lebiere's ACT-R system (1997), Norwich's entropy theory of perception (1993), and Link's wave theory of psychophysics (1992). A ®nal point that needs to be reiterated here is that theories and models, however formal, serve as a guide to the collection, analysis, and interpretation of empirical data. There are no such things as data divorced from a theoretical framework, just as there is no such thing as an object divorced from its inertial framework (Newton). Even an informal, ``commonsense'' framework gives meaning to data collected within it, which is only to emphasize the importance of being explicit about our theories and models. Although the mere selection of phenomena to observe reveals what the accompanying theory is about, it is not always enough to reveal the assumed relationships between those phenomena and an informal, implicit theory or model. Thus, to take full advantage of Weinberg's principle of complementarity (chap. 7), let us always state our guiding theories explicitly and precisely. 8.2

Statical Models

Most theories and models in psychology are statical models. As mentioned in chapter 1, a statical model assumes, at least implicitly, that the relevant state of the modeled system remains constant, and speci®es the

Dynamical versus Statical Models

63

relations between the various state variables of that system. If the model is formal, there is an equation of state that relates the measured values of the variables for some particular state. For example, either Fechner's or Stevens's psychophysical law is an equation of state. Let us revisit Stevens's law (eq. 1.3): R ˆ cS m ; which states that, whenever the psychophysical system (experimenter, stimulus, subject, response; see Ward 1992) has a particular value of the experimenter's stimulus, S, the value of the subject's response, R, will be cS m (with some error, of course). Although equation 1.3 is the simplest kind of equation of state, relating just two variables of a system, much can be done with it (see, for example, Geschieder 1997; Marks 1974). Music halls are designed according to its predictions of responses to the intensity of the music at various places in the halls, lawsuits are launched against polluters regarding the smelliness or loudness of their byproducts, and theories of sensory system function are tested noninvasively by measuring the parameters, especially m, of this equation under various conditions. What is perhaps most important, equation 1.3 summarizes fundamental relationships between important concepts; indeed, in some cases, it de®nes the concepts themselves. Stevens (1975) interpreted the exponent m to be the operating characteristic of the sensory transducer: di¨erent values of m measured for di¨erent sensory modalities (e.g., brightness in vision and loudness in hearing) re¯ected the di¨erent energy transduction principles of the various sensory modalities. Possibly more familiar to most readers will be the many statical models of the human memory system that have appeared over the past 150 years. Enduring theoretical distinctions are those between short-term (or primary) and long-term (or secondary) memory systems and between episodic and semantic memory systems; a more recent popular model distinguishes between implicit and explicit memory systems. These informal models each de®ne two conceptually separate memory systems that vary on several key properties. For example, the explicit memory system is said to control memory performance that is ``initiated and guided by a conscious plan or by an intention to recollect speci®c prior episodes'' (Graf 1994, 682), whereas the implicit memory system is involved when some prior episode in¯uences present behavior in the absence of any such conscious plan or intention to recollect that episode. This leads to several empirical practices, such as asking subjects explicitly to remember in tests of explicit recall but giving them various priming tasks, such as stem

64

Chapter 8

completion, to test implicit memory. As proposed by Mandler (1980), implicit and explicit memory are said to be related to the underlying psychological processes of integration and elaboration. In this view (see also Graf 1994), implicit memory is the result of integration processes that make connections between the various features of perceptual objects during their perceptual processing, whereas explicit memory results from the elaborative processing that relates di¨erent perceptual objects in terms of their meaning. Note that although memory is inherently a temporal phenomenon (see chap. 5), such models emphasize the relationships between variables outside of time. For example, much is made in the memory literature of the observation that amnesiac patients often show substantially more implicit memory, sometimes near ``normal,'' for past episodes than they do explicit memory (indeed, amnesia is de®ned by a lack of explicit memory). In other words, the (informal) state equation is as follows: amnesiac state implies low (relative to normal) explicit memory but high implicit memory, whereas nonamnesiac state implies high explicit memory and high implicit memory. 8.3

Dynamical Models

A dynamical model is concerned with a succession of states, that is with change, and thus with time because that is the dimension associated with change. Every dynamical model has time as a variable, although it is often represented implicitly. The prototypical dynamical model is a differential equation, such as the following simple linear one: dx ˆ at; dt

…8:1†

where dx=dt, the derivative of x with respect to t, represents the rate of change of the variable x with respect to time, t. The dx and the dt, called ``di¨erentials'' (after which di¨erential equations are named), refer to in®nitesimal changes of x in in®nitesimal intervals of time. The equation describes the limit of change in the value of the state variable x as the change in t (time interval) becomes very small. Equation 8.1 states that the rate of change increases linearly with t, beginning at the value 0, when t ˆ 0, that is, a
0 ˆ 0. With time measured in seconds, if a ˆ 0:001, then 1,000 seconds after t ˆ 0, the variable x is changing at a rate of 1 unit per second, whereas at t ˆ 10; 000, it is changing at a rate of 10 units per second. Thus the rate of change increases with time. What does this mean about the behavior of the state variable x? We might suspect that, because

Dynamical versus Statical Models

65

the rate of change just keeps on increasing, meaning that x keeps changing faster and faster forever, x itself would grow forever (or until something broke down). To ®nd out for certain, we solve equation 8.1, which involves multiplying both sides of the equation by dt and integrating each side with respect to the variable in the di¨erential on that side (integration in calculus is the inverse of di¨erentiation). Here the solution is an explicit function of time because t appears on the right-hand side of the di¨erential equation. When we solve equation 8.1 in this way, we obtain as one of the solutions xˆa

t2 : 2

…8:2†

(For the purists, I assumed that the di¨erence of the two constants of integration equals zero.) Now, beginning with x ˆ 0 at t ˆ 0, it should be clear that x will grow forever on an ever-steepening curve, just as we suspected. This may be a useful model of the national debt, or love in a lover's dream, but it is not going to be too useful in dynamical cognitive science. Most useful models in dynamics have more interesting, and more limited, behavior. In particular, feedback often limits the growth of a state variable, as we will see shortly. Let us ®rst consider another di¨erential equation, however: dx ˆ ax: dt

…8:3†

Because t does not appear on the right-hand side of the equation, the solution of this di¨erential equation is an implicit function of time. To solve it, we need to rewrite the equation as dx=ax ˆ dt and then integrate as before. The result is (again assuming that the di¨erence of the constants of integration is zero) x ˆ e at ;

…8:4†

as is well known in the literature of di¨erential equations. In this exponential growth equation, if a > 0, then there is again growth without limit, whereas if a < 0, there is a decline toward an asymptote of zero. If a ˆ 0, then x ˆ 1 for all time, that is, the system does not change with time. There are many other, more useful, di¨erential equations, to be sure, and most are functions of more than one variable, although they are usually not easy to solve, especially when they are nonlinear. (For those interested in actually doing this kind of modeling, Morrison 1991 provides summaries of many of the important techniques.) It is often possi-

66

Chapter 8

ble, however, to ®nd an approximation to di¨erential equations in the form of di¨erence equations, and work with these on a computer, as we will see shortly. Without feedback, a dynamical model is boring, whereas, with it, the model becomes both useful and fascinating. Let us add some feedback …ÿbx 2 † to equation 8.3: dx ˆ ax ÿ bx 2 dt

…8:5†

Notice that ÿbx 2 is a negative term and will decrease the rate of change of x more and more as x gets larger and larger, o¨setting the tendency of the rate of change to increase linearly with x according to the ®rst term. Solving Equation 8.5 results in another implicit function of time, xˆ

c ; 1 ‡ …c=x0 ÿ 1† exp‰ÿa…t ÿ t0 †Š

…8:6†

where c ˆ a=b. Figure 8.1 shows a graph of this logistic equation. When 0 < x0 < a=b, a > 0 and b > 0, then x approaches x0 ‡ a=b as t > t0 increases. Starting at t0 , growth is exponential for a while but then, as the feedback begins to make itself felt, slows down, leading to an asymptote of the variable at x0 ‡ a=b. Equation 8.6 is more useful than equation 8.4, having served for example, to model the consumption of a nonrenewable resource, such as petroleum (see Meadows et al. 1972). Indeed, its appli-

Figure 8.1 Graph of the logistic equation (eq. 8.6) for a ˆ 0:1, b ˆ 1.

Dynamical versus Statical Models

67

cation in this area predicted that domestic oil production in the United States would decline as the available oil was discovered, produced, and consumed, until there was little left. Similar models might be usefully applied in cognitive science to the consumption of limited cognitive resources, such as memory, attention, or processing e¨ort. In this case, however, the temporal unfolding might be over seconds, minutes, or hours, rather than over years. There are two other important ways of setting up dynamical models. One is to use a di¨erence equation, such as the logistic di¨erence equation xn‡1 ˆ axn ÿ axn2 ;

…8:7†

whose resemblance to equation 8.5 is apparent (though it behaves altogether di¨erently). A di¨erence equation assumes that time changes in equal and discrete jumps, so that the index n is taken to be integers representing equally spaced moments of time. Di¨erence equations are closely related to di¨erential equations, and often di¨erential equation models can be turned into a di¨erence equation models and studied by computer simulation, an approach much in vogue today. The relation to di¨erential equations can be made clearer by subtracting xn from both sides of equation 8.7 and then writing the left side, xn‡1 ÿ xn , as Dx. Dividing both sides by Dt ˆ 1 (by assumption), which leaves the right side unchanged, and suppressing the subscript n on the right, we obtain Dx ˆ x…a ÿ 1† ÿ ax 2 ; Dt

…8:8†

which looks very much like di¨erential equation 8.4, with the ®nite differences Dx and Dt replacing the di¨erentials dx and dt. I will discuss the striking behavior of the logistic di¨erence equation 8.7 in chapter 24 and show how it can be used as a model for memory in chapter 26. (For a very good practical introduction to the details of working with di¨erence equations, see Goldberg 1986, which also includes examples of models in psychology and other ®elds.) The other approach to setting up a dynamical model is simply to add a time index to a statical model. For example, equation 1.3 can be turned into a dynamical model simply by adding a time index to the variables, in this case a discrete trials index, i: Ri ˆ cSim ;

…8:9†

which describes the response on the ith trial as a function of the stimulus intensity on that trial. Now the model states that all of the trial-to-trial

68

Chapter 8

variation in psychophysical judgment responses can be accounted for by the variation in nominal stimulus intensity (controlled by the experimenter) from trial to trial. Because this is not the case in any published scaling experiments, an error term must be added to equation 8.9 to account for the residual variance. Now the question becomes, is the error simply random, or is it correlated with some other variables, in particular with previous values of Ri or Si , say Riÿ1 or Siÿ1 ? The model of equation 8.9 can usefully include more terms and even if it does the residual variance might still retain some structure. Chapter 11 summarizes the dynamical analysis of psychophysical judgment using equation 8.9 as a starting point. DeCarlo and Cross 1990 presents the detailed development of one of the resulting models of psychophysical judgment and also describes this method of developing a dynamical model in a way applicable to other modeling problems. 8.4

Why We Need Both Statical and Dynamical Models

Statical models seem to be more popular, and are arguably more useful, in biology, psychology, sociology, and so forth, than they are in physics and economics, at least partly because in the former ®elds the fundamental concepts depend more on the classi®cation of phenomena. In other words, specifying the fundamental phenomena requires recognizing that a particular event falls into a particular category (Rapoport 1972). For example, in psychophysics, the decision to study sensory magnitude as a function of the stimulus intensity requires that ``sensory magnitude'' be given meaning. Fechner (e.g., 1860) spent a great deal of time and e¨ort doing just this; psychophysicists have continued his e¨ort for over 140 years, making invaluable use of the statical models of Fechner's and Stevens's laws as they have (cf. Ward 1990). Similarly, the study of human memory has involved repeated e¨orts to develop concepts that capture the important memory phenomena in ways that allow us to discover reliable empirical regularities. The distinctions mentioned earlier, and the models they have spawned, have led to impressive advances in our understanding of how behavior depends on previous experience. Thus, without useful statical models in these ®elds, it is doubtful that scientists could agree on the relevant phenomena of study. Although a dynamical model can always be built as long as a time series of events can be obtained for study, the meaning of the events used to inform the dynamical model needs to be anchored in some larger conceptual framework. A highly con®rmed statical model can be a useful anchor. More-

Dynamical versus Statical Models

69

over, when studying a ®eld in which there are many useful statical models, the body of knowledge they summarize both informs any future e¨orts at model building and concept discovery, and provides a fertile ground for testing the generality and accuracy of dynamical models. Statical models produced by collapsing dynamical models over time must agree with other, highly con®rmed statical models or the original dynamical models are suspect. For example, any dynamical model of psychophysical judgment must in some way or other collapse to a statical psychophysical law (e.g., eq. 1.3; see Gregson 1988). Thus, consistent with Weinberg's principle of complementarity, dynamical models complement existing and future statical modelsÐthey do not replace them.

This page intentionally left blank

Chapter 9 Dynamical and Structural Models

One of the more remarkable aspects of the connectivity of the cerebral cortex is the fact that when a given area A sends an output to area B, area B sends a return output back to area A. . . . This is such a ubiquitous arrangement that there are only a few exceptions to it. ÐSemir Zeki, A Vision of the Brain These topological features are at times the most essential ones in the descriptions of many systems. ÐAnatol Rapoport, ``The Uses of Mathematical Isomorphism in General Systems Theory''

9.1

Structural Models

I argued in chapter 8 that statical models are useful to dynamical modeling because they provide a framework within which dynamical models can be constructed and tested. Of course they are also useful in their own right because they de®ne, categorize, and often make sense of the elementary phenomena with which a science is concerned. In both of these roles, statical models often take the form of what I call ``structural models.'' That is, they achieve their goal by specifying a set of elements and describing the interrelations between these elements, emphasizing which elements are connected to which other elements (and can thus a¨ect them causally). This emphasis on connectedness of elements is central to the role of structural models in dynamical cognitive science because it provides signi®cant constraints on the form of any dynamical model describing the time evolution of the structurally modeled system. A compelling example of the importance of such constraints comes from a comparison of various structural models of the brain. Descartes modeled the brain as a pump that distributed the humors, via the nerves, to the muscles and so caused behavior, much as a water pump distributes

72

Chapter 9

water through tubes to drive fountains. Clearly, any dynamical model of the brain based on Descartes's theory would involve hydraulic theory, and its basic elements would be pump chambers, tubes, and ¯uids. In contrast, the telephone switchboard model of the brain, in vogue after the electrochemical nature of nerve conduction was discovered and telephones invented, would require wires and switches and would involve the ¯ow of electrons through those wires. Finally, the more modern models of the brain, based on computer metaphors, require elements such as ¯ip-¯ops organized in registers, again with wires communicating between them. In these models, however, the ¯owing substance is digitally encoded information (in the form of electrical pulses). Although dynamical models derived within these three frameworks would undoubtedly share some aspects, they would di¨er from one another more often than not. For example, if a tube ®lled with ¯uid is cut, the ¯uid can drain out of any connected part of the system, causing total breakdown (as when an artery is cut and all of the blood drains from a body). In contrast, if a wire is cut, electrons can no longer ¯ow from one terminus of the wire to the other, but electron ¯ow is still possible in the other wires of the circuit (however altered it might be according to the laws of electrical circuits). Finally, if an information channel is cut, clearly information can no longer pass through that particular channel, but, rerouted, it might very well get to its destination by another pathway, and the system's functioning might be a¨ected hardly at all. 9.2

Graph Theory

An excellent formal system within which to model system structure, graph theory involves a graph, de®ned as a set of points (vertices or nodes) in space and a set of interconnecting lines (edges) between them; for each pair of vertices, there either is or is not a connecting line. If the edges have an arrowlike direction associated with them, then the graph is said to be ``directed''; each vertex has an in-degree, the number of edges directed toward that vertex, and an out-degree, the number of edges directed away from that vertex. Figure 9.1 shows an example of a directed graph, with the edge direction speci®ed by the arrowhead on each edge. Finite state machines (see ®g. 5.1), electrical circuits, and perhaps even brain circuits are all directed graphs, whose important function is to specify what is connected to what. A path between two vertices is a list of the vertices that have to be traversed to get from one to the other; DAC is a path between vertices D and C in ®gure 9.1. A path is simple if, as in

Dynamical and Structural Models

73

Figure 9.1 Directed graph.

DAC, no vertex in the list is repeated. A path such as ACBA, because its beginning and ending vertex are the same, is called a ``circuit'' or ``cycle,'' (a circuit is simple except for the ®rst and last vertices). A graph with no cycles is called a ``tree.'' Favorite data structures in computer science, trees have also been used for structural models in psychology (e.g., Sattath and Tversky 1977). A graph is connected if there is a path from every vertex to every other vertex in the graph. The graph in ®gure 9.1 is not connected: there is no path connecting C to D, among other pairs of vertices. If a graph with N vertices has fewer than N ÿ 1 edges, it cannot be connected; if it has more than N ÿ 1 edges, it must have a cycle and thus cannot be a tree. Although a tree has exactly N ÿ 1 edges, not all graphs with exactly N ÿ 1 edges are trees. If there are weights on the edges, as in ®gure 5.1, where they represent state transition probabilities, the graph is a weighted graph. Finite state machines are thus directed, weighted graphs. Because graph theory is an enormous topic in mathematics, I can only, as before, touch on a few less enormous topics relevant to it (for a basic text, see Berge 1985). The computer science literature contains a huge range of graph theory applications, including many relating to cognitive science, for example, expert systems (see Cowell et al. 1999) and directed weighted graphs used as models of stochastic systems (for an excellent detailed account, see Lauritzen 1996). Many theorems have been proven about various types of graphs, especially about the simplest ones. Many of these concern methods to ascertain whether and how many paths or cycles exist in a certain type of graph, or what might be the most e½cient way to get from one vertex to another. The celebrated traveling salesman problem, in which all of the vertices in a graph must be traversed at least once, by the shortest possible route, is an example of a particularly fruitful

74

Chapter 9

graph theory problem. These static graphs might be used directly as the basis for a dynamical model concerning the temporal unfolding of the system state de®ned by the vertices. Such a model has been suggested for the cognitive neuroscience of attention, in which di¨erent, connected brain areas ``light up'' (show activity above the background level) successively as a cognitive task is performed (e.g., Posner and Raichle 1994). Of course graphs can change. The main category of dynamical graph studied is the random graph, where links ``grow'' among the initially de®ned set of vertices, and where a probability distribution over the vertices associated with the growth of the links between them provides constraints on the emerging structure. For a given set of probability distributions, there are many possible, di¨erent realizations of the random graph so de®ned. The relevant theorems of such a graph de®ne the probability distributions for whether the graph is connected, how many indegrees its vertices have, how many paths there are from one vertex to another, and so forth. One common application of such graphs is that of predicting the extent and lifetime of disease epidemics. Cognitive science applications might include understanding the spread of information in a brain or in a culture. Although much of the sophisticated mathematical analysis of directed graphs can be done only on graphs that have no cycles, such as trees, real systems always have cycles, as do most abstractions of real systems, so that the directed graphs representing them should also have cycles. This situation presents a dilemma for mathematical analysis. One solution, suggested to me by Priscilla Greenwood (personal communication), involves a particular way of introducing dynamics to the graphical structure. It can be illustrated by a planar graph, such as in ®gure 9.1, which has a cycle. The idea is to add a unidirectional time dimension perpendicular to the spatial dimensions (think of it as rising vertically from the page for ®gure 9.1), and then to replicate copies of the graph along the time dimension. Each copy of the graph represents one instant in time during which only one edge is ``active'' (represented by, say, a solid line), whereas the remaining edges are passive (represented by, say, dotted lines). This scheme transforms cycles in space into helices (or spirals) in space-time. For example, beginning at vertex A at time t1 the cycle ACBA in ®gure 9.1 becomes the helix (A; t1 ) (C; t2 ) (B; t3 ) (A; t4 ). Because of the di¨erence in time coordinates between (A; t1 ) and (A; t4 ), the helix never gets back to the origin, there are no cycles, and mathematical analysis can proceed on the augmented graph. This is just one illustration of how the introduction of a time dimension can make a model more useful and the

Dynamical and Structural Models

75

modeler's task easier in some ways (see also Weinberg 1975; chap. 6). Of course the analysis of helices is routine in molecular biology and, interestingly, also in music theory, where musical pitch takes the form of a helix. Directed graphs are already used by many cognitive scientists as models of system structure. The example of such a model shown in ®gure 9.2 is a ¯ow diagram, albeit a very complicated one. Of course such models are not analyzed to determine whether they are connected (they usually are not), or how many ``shortest'' paths there are between particular nodes, or really for any of the myriad other mathematical properties of directed graphs. In these ¯ow diagrams, it is the graph's local properties, such as the content of the nodes and the presence or absence of particular edges, that dominate its analysis. The model is really a ``su½ciency'' demonstration, showing how the properties of a particular empirical phenomenon, such as the experience of a visual image in the case of ®gure 9.2, could arise within the structure of the functional nodes depicted. Such a ¯ow diagram also could be dynamicized by adding a time dimension and allowing only one active edge per frame. I can imagine a quite complicated time evolution in such a complicated graph, one that might even resemble some of the information processing done in support of everyday human behavior (given that this system is supposed to represent the human mind). Moreover, because the graph in ®gure 9.2 certainly has cycles, such treatment would allow some of the more global, graphtheoretic properties of the model to be explored, perhaps with interesting implications for understanding the behavior the graph represents. Finally, it is likely that explanation of human behavior would require that several edges be simultaneously active. As long as the active edges do not form any cycles within a single frame, the mathematical analysis should not be compromised. The result would be a dynamical model with several helices developing, and possibly intertwining, over the same time period. A model like this, with one helix emphasized and the others relatively suppressed, has been suggested, in principle, as a description of human consciousness (Dennett 1991). It is also possible to see the brain itself as a directed, weighted graph, as mentioned earlier and as hinted at in the ®rst epigraph at the beginning of this chapter. Indeed, topological maps of brain connections are commonplace in neuroscience. A famous example is the map of the myriad interconnections (edges) between anatomically distinct parts (vertices) of the macaque monkey visual system. Figure 9.3 shows a modern version by van Essen and Deyoe (1995). Although portrayed as an undirected

76 Chapter 9

Figure 9.2 Directed graph of the human mind. (From Kosslyn and Koenig 1992. ( 1992 by Stephen M. Kosslyn and Olivier Koenig. Reprinted with permission of The Free Press, a Division of Simon & Schuster, Inc.)

Dynamical and Structural Models

77

Figure 9.3 Graph of the monkey visual system. (From van Essen and Deyoe 1995. Reprinted by permission of The MIT Press.)

78

Chapter 9

graph, most if not all of the edges actually are two-way, or rather two di¨erent edges, one in each direction (Zeki 1993). Although the accurate depiction of the complete form of such a model is a distant goal of the ®eld of neuroscience, useful submodels might already be possible in the spirit of ®gure 9.3 and of the work of Posner and Raichle (1994). Such structural models would depict several anatomically connected, functionally related brain areas, and would specify the order of functional activity in performance of a particular behavior by depicting the series of active edges (and nodes) along the time axis. Several such models could be juxtaposed, with appropriate connections, to expand the model's applicability, and the possibility of several edges being simultaneously active could be introduced. Again, the temporal unfolding of such a model would be complicated, and would be a starting point for the development of a dynamical model of brain activity that would admit of mathematical analysis to discover the global properties of the structure and their implications for brain function. A ®nal relevant topic in graph theory is that of the scale of the graph. Whereas a simple system can be represented by a simple planar graph such as in ®gure 9.1, or by a more complicated, time-evolving graph such as those just discussed, both types consisting of just a single scale level, many complex systems, with multilevel groupings of vertices, whose edges connect the nodes within a group and the various groups of nodes, must be represented by multiscale graphs such as the graph in ®gure 9.2. These multiscale graphs, though much more di½cult to analyze mathematically, are also more likely to be realistic models of actual cognitive systems. (Several examples, especially those producing ``colored'' noise, will be discussed in later chapters.) 9.3

Interplay between Dynamical and Structural Models

As just discussed, the dynamics of a system are both informed and constrained by the structure within which they operate. Dynamics also can create structure, however, such as a heat cell when a ¯uid is heated from below, which then sustains itself. Here the original structure of the ¯uid, which consisted of loosely interconnected molecules moving randomly with respect to one another, changes when heat is added in a particular way to a much more tightly bound set of molecular motions in particular, spatially distributed directions. In the case of a complex system such as the brain, there are important distinctions to be made between statical structure (e.g., sensory pathways), dynamical structure (e.g., synaptic

Dynamical and Structural Models

79

plasticity that creates changing connections between neurons), dynamically created structure (e.g., momentarily active cell assemblies, conceptually in the same category as a heat cell), and pure dynamics (e.g., a stochastic process or a ®nite state machine). We can always ®nd revealing structure, even in pure dynamics (e.g., a stationary law in a stochastic process; see chap. 12). Thus we must face the prospect that a dynamical model alone, even if deliberately created without a structural basis, still introduces structural aspects to its analysis. We must be prepared to recognize and deal with this structure, for it often forms an important product of the dynamical analysis. Indeed, some authors consider the geometric ¯avor of nonlinear dynamical analysis to be its most important, and in many cases its only, usable product (e.g., Kellert 1993). Finally, there is the prospect of building dynamical models that implement realistic structural assumptions even as they explore the consequences of changes in structure, whether of the brain or of more abstract cognitive systems. Neural network models are one promising example of this possibility, although, to portray ongoing dynamics, they would have to be developed somewhat di¨erently from how they are today. Such models are directed, weighted graphs at one level, although their processing power vastly exceeds such a simple structure. Chapter 34 reveals the tip of the neural network iceberg in the context of dynamical models.

This page intentionally left blank

Chapter 10 Deterministic versus Stochastic Dynamical Models

A linear universe is a physical impossibility. . . . Even numerical methods do not cope with all the di½culties of nonlinear equations. . . . Statistical modeling must be applied at some level or stage of development. ÐFoster Morrison, The Art of Modeling Dynamic Systems

10.1

Deterministic Models

All of the models discussed thus far have been deterministic, which simply means that they contain no random elements. And what is random? I explore this very di½cult and important question in several contexts (especially in chaps. 11 and 22). The answer is elusive; indeed, as Mark Kac (1983, 406) has stated: ``From the purely operational point of view the concept of randomness is so elusive as to cease to be viable.'' He meant that it is impossible to tell whether a particular sequence, say of heads and tails produced by coin tosses, was produced by actual coin tosses or by an arithmetical procedure designed to mimic the outcomes of coin tosses. In this sense, a deterministic, completely predictable, arithmetical process behaves in the same way as something presumed to be random and unpredictable (unless the coin is severely biased). To take this even further, it is possible to create a computer algorithm (a deterministic one that produces the same sequence every time it is run) that will mimic any ``random'' process whatever (see also Kac 1984). Thus, when I say the model ``contains no random elements,'' I mean to say that it is a simple one-to-one function, that the values of the state variables completely determine a system's state, and that there is no variability of behavior of the system for a given set of state variable values. For example, in the statical model of equation 1.3, R ˆ cS m , each value of S gives rise to one and only one value of R, always the same one regardless of

82

Chapter 10

other conditions. Similarly, even the dynamical version of that model (equation 8.9), Ri ˆ cSim , has the same property. On each occasion, i, there is one value of S and one value of R. Whenever the S is the same as some other S, the R is also the same. There is no error variability; the value of S completely determines the value of R. Of course these are ideal, theoretical models, although the model is supposed to express some underlying ``truth'' that obtains in the real world. However, tests of such theoretical statements always involve the ``real'' real world and all of its myriad complexity. We can never isolate our system from all other in¯uences. Thus, when we set up an experiment in psychophysical scaling, say, and present a particular S to our subjects on several di¨erent occasions, we ®nd that the same subject often responds with di¨erent Rs on those di¨erent occasions. There is variability in R even though the value of S was nominally (not actuallyÐthat is also impossible to control exactly) the same. Moreover, di¨erent subjects also respond to the same S with di¨erent Rs even if only presented once. In what sense, then, could we say such a deterministic model ``holds''? We assert that the model holds ``on average,'' that it holds for the central tendency of the di¨erent values of R observed for each value of S. We accept the di½culty of isolating our experimental system from all other in¯uences and simply relegate the e¨ects of those in¯uences, giving rise to the variability in R, to a ``noise'' term. E¨ectively, we divide our models into two classes: theoretical and empirical. The theoretical model is what we want to assert, whereas the empirical model is realistically what we have available. All empirical models contain a noise term, re¯ecting the uncontrolled in¯uences in our experiment or observation. Theoretical models often do not contain a noise term because we are uninterested in ``other in¯uences'' we cannot control. Even when the theoretical models do contain a noise term, however, it is almost never a term that attributes any structure to the noise, particularly temporal structure. It is almost always only the amount of noise that we model, not its temporal distribution (again, there are signi®cant exceptions). Thus far we have considered only the simplest deterministic models. Because they change either not at all or only very slowly relative to the timescales of interest, all statical models are called ``type 0'' by Morrison (1991). The simple dynamical models he calls ``type I'': here, because of the precision with which they can be speci®ed, theoretical models are clearly distinguished from their empirical counterparts. When we move into the realm of di¨erential and di¨erence equations things become a bit less clear, although there is still a class of models among these that yields

Deterministic versus Stochastic Dynamical Models

83

precise solutions. These are called ``class I, type I'' systems by Morrison (1991). They include di¨erential equations that can be solved precisely, without error, such as equation 8.1 (with solution eq. 8.2). We run into trouble almost immediately, however, when we encounter di¨erential equations whose solutions involve unavoidable error, ``class II, type I'' systems such as equation 8.3 (solution eq. 8.4). The solution is an exponential equation; because the number e is irrational (its exact value can never be known, only approximated with increasing accuracy), there will always be some rounding or truncation error involved in its evaluation. Although this will often be trivial, sometimes it will not (see Morrison 1991, 178). The problem of error becomes more severe for what Morrison (1991) calls ``type II'' systems, used when data are very precise or abundant, so that the data depart from any type I model. Here a special technique creates a series of models and uses power series expansions to account for the departure of the real system from the ideal model, with each created model providing somewhat better approximation to the data until the accuracy is su½cient for the purpose at hand. Although as Morrison states and demonstrates, these highly intricate and di½cult methods are of dubious value today because of the availability of numerical simulations on computers, nonetheless, they illustrate that, even for a relatively simple deterministic system, small perturbations (such as those caused by the planets of our solar system acting gravitationally on each other) can be important and can complicate the modeling e¨ort. ``Chaotic'' or ``type III'' systems pose an additional problem: they are di¨erential or di¨erence equations whose behavior is so complex it actually seems random in many ways (see chaps. 23 and 24). Such equations do not have to be complicated to behave in a complicated way. The logistic di¨erence equation, (e.g., eq. 8.7) discussed repeatedly in this book and in every book dealing with chaos theory, is a perfect example. Exceedingly simple in its geometric form (its nondynamical map is a parabola), it behaves wildly and cannot be predicted in principle. On the other hand, many systems of nonlinear di¨erential equations also behave wildly and have geometric forms that are more complicated and di½cult to grasp. As we will see in chapter 24, it is di½cult to distinguish chaos from randomness, although many people like the idea of the underlying determinism in chaotic models (notice there is no ``random variable'' in equation 8.7). Although computer simulation techniques have been a great help, we are still often faced with ``Bonini's paradox,'' where the behavior of a computer simulation model is as di½cult to understand as

84

Chapter 10

that of the empirical system that inspired it. Various techniques exist to analyze these models, including statistical ones. Their importance here is that they emphasize that even deterministic models can behave in ways that are di½cult to distinguish from randomness. Indeed, as soon as we move away from the simple, statical, idealized model represented by Stevens's law (and many other psychological theories), we encounter randomness. Although this randomness shares many characteristics with the more usual random error that permeates our data and that sometimes is the basis for powerful models, it also contains di¨erent, exotic elements that arise from temporal unfolding. 10.2

Stochastic Models

A stochastic model is one in which randomness plays some role, often a profound role inherent to the phenomenon modeled. A simple dynamical empirical model of psychophysical judgment, such as ^ log Si ‡ log c^ ‡ ei ; log Ri ˆ m

…10:1†

is stochastic. In this model the ei term is the random variable, which represents a series of samples from some probability distribution, usually the normal distribution with mean zero and some nonzero standard deviation. In the empirical model, the random variable is the ``noise'' termÐ it is simply there to represent the variability that cannot be captured by the model. When we move into the theoretical domain, however, because the ``noise'' is not taken to be important theoretically and constitutes a considerable nuisance during theoretical calculations or derivations, the random variable is commonly dropped. One of the major themes of this book is that this practice has concealed some extremely interesting phenomena from us. That neglected noise often has temporal structure, and when it does, it deserves to be modeled. In particular, the temporal relationships buried in the noise can reveal important aspects of the process under study (see chaps. 14 through 19). There are several domains in psychology where formal modeling has been quite well developed and successfully applied to empirical data, among them learning, measurement, perception, attention, and psychophysics. The ®eld of mathematical psychology has become quite sophisticated, with a wide variety of modeling approaches, although it continues to have less impact on mainstream psychology than it deserve to. Importantly, stochastic theoretical models have played a large role in mathematical psychology; indeed, entire books have been written on the subject,

Deterministic versus Stochastic Dynamical Models

85

including Green and Swets 1966, a classic on signal detection theory (the sina qua non of stochastic statical models); the excellent Townsend and Ashby 1983; and Luce 1986 which concentrates on response times, as reaction times are often called. These books, and many others, describe myriad stochastic theoretical models, including some dynamical ones. For example, Luce 1986 describes a model by Pacut (1978), where information about a sensory signal, whose appearance is to be reacted to as quickly as possible (simple reaction time), accumulates until it exceeds some response criterion, at which time a response is made. Only one of the many models proposed to describe simple reaction time, the Pacut model assumes that the accumulation of information about the signal is perturbed by ``additive random noise.'' (Luce, 1986, 154). In this model, A…t† represents the total information accumulated by time t; and a response is made when A…t† > C…t†, where C…t†, the criterion, is a normally distributed random variable. A…t† is in¯uenced linearly by S…t†, the signal amplitude, which is taken to be 0 for t < 0 and 1 for t b 0, and additively by w…t†, the noise, again normally distributed with mean zero and variance proportional to elapsed time from signal onset. The model can be written as a di¨erential equation: dA…t† dw…t† ˆ ÿb‰A…t† ÿ aS…t†Š ‡ ; dt dt

…10:2†

which states that the momentary change in accumulated information about the signal is proportional both to the signal amplitude and to the current amount of information, plus some noise. But for the noise, equation 10.2 would be a simple linear di¨erential equation; the deterministic part would be a ``type I'' model. The addition of the noise complicates matters however. The solution to equation 10.2, for A…0† ˆ A0 , where A0 is also a normally distributed random variable, is   …t ÿbt bx A0 ÿ a=t ‡ e dw…x† ; …10:3† A…t† ˆ a ‡ e 0

which states that at any time, t, the accumulated information is a random variable composed from a constant, a, representing the coupling to the signal, plus a complicated, random term representing both the (random) starting amount of information, A0 , and the weighted integral of the noise over the time since signal onset. This random walk from a random starting place proved to be too complicated to apply directly (Luce 1986), although a special case does ®t some reaction time data fairly well, and a discrete version would not be di½cult to simulate. Notice that this model

86

Chapter 10

describes what happens over time on a single trial of a reaction time experiment. Such models are typically evaluated by deriving their predictions about the variation of expected values of reaction times with causal factors, such as signal amplitude, and about the (usually somewhat asymmetrical) form of reaction time distributions, in both cases collapsing across all of the trials on which reaction time is measured under a given set of conditions. Only a single scale of time, that of a single reaction time trial amounting to perhaps several hundred milliseconds, is modeled. Even though there is no attempt to model either longer timescales (e.g., an entire experiment) or shorter ones (e.g., neuronal conduction times) nonetheless, the model is already ``too complicated.'' Another approach to stochastic modeling in psychology is to represent response variability directly, usually by making the dependent variable of the model the probability of a response. All signal detection theory models are of this character, as are most of the very general psychophysical models described by Falmagne (1985). In one such statical model of psychophysical judgment, p…yes† ˆ ‰1 ‡ eÿ…I ÿm†=y Šÿ1 ;

…10:4†

the probability of saying, ``Yes, I see it,'' to the presentation of a faint ¯ash of light is modeled as a logistic function of the light intensity, I, with m representing the absolute threshold for light under the given conditions, and y representing the steepness of the change from never seeing the light to always seeing the light (the ``fuzziness'' of the threshold). This function has a similar graph to that of ®gure 8.1, except that here the abscissa is stimulus intensity rather than time. In such a model, the decision process is considered to be hidden, so that only probabilistic predictions can be made from the model; the outcome of any single trial in an experiment is unpredictable. This situation is reminiscent of quantum theory (see chap. 13), in which randomness is considered to be necessary, and not simply the result of ignorance. Of course, such a model could be made dynamical by introducing time, and made even noisier by adding a noise term. Alternatively, a dynamical model could be constructed from the start, in which the temporal unfolding of the decision process would give rise to the variability in the psychophysical decisions made in such a way that the model of equation 10.4 would describe the expectation of the dynamical model's outputs as a function of light intensity. Thus, for many di¨erent reasons and in many di¨erent ways, cognitive science must utilize, and has utilized, stochastic models. Many sophisticated stochastic models have already been proposed, although most are

Deterministic versus Stochastic Dynamical Models

87

either statical, or dynamical over only a limited range. Nonetheless, the techniques used in these models have proved to be highly e¨ective, and will continue to be so when applied in a wider dynamical context. Still more sophisticated techniques are available in mathematics and in physics, as the material discussed in chapters 12 and 13 reveals; such techniques will come into wider use as dynamical cognitive science matures. 10.3

Do We Need Both?

Deterministic models capture the basic, nonstochastic processes of a system, addressing central tendencies in an elegant (though not always simple) way, whereas stochastic models address variability directly, albeit sometimes only by averaging over it or putting it into a catchall random variable (noise) term. Importantly, the two types of models can lead to di¨erent outcomes even when applied to the same data. This is perhaps most clearly demonstrated in the context of psychophysical scaling, where S. S. Stevens (1975) pointed to the con¯ict between the poikilitic (based on variability) function called ``Fechner's law,'' R ˆ a ln…S=S0 †, and Steven's own psychophysical power law, R ˆ cS m , based on central tendencies of direct scaling judgments. Because these two functions can be ®tted to di¨erent statistics of the same judgments, they actually reveal two different aspects of those judgments, another example of the ubiquity of Weinberg's principle of complementarity (chap. 7). The same possibilities exist in dynamical models. In the dynamical context, stochastic models account for or describe the structure of ¯uctuations over time, whereas deterministic models describe trends. Useful models sometimes have both deterministic and stochastic parts (see chaps. 11, 12, 30). And, as pointed out above and in other places (e.g., Morrison 1991), the distinction between variability and trend blurs when the deterministic model is nonlinear, as in chaos (see chap. 24).

This page intentionally left blank

Chapter 11 Linear Time Series Analysis

ARIMA models are not arbitrarily ®t to data, but rather are built empirically. ÐRichard McCleary and Richard A. Hay, Jr., Applied Time Series Analysis for the Social Sciences

The vast literature on linear time series analysis can only be woefully adumbrated in fewer than ten pages. Indeed, it is only because the literature is so vast and the number of helpful works so numerous that I can get away with this. The most useful of the helpful works, in my opinion, is McCleary and Hay 1980, quoted above, on which the ®rst two sections of this chapter are mostly based. Another good book is Gregson 1983, which argues forcefully for a change in emphasis in psychology from statical to dynamical modeling, although those without excellent mathematical background will ®nd it hard going. A third, for those preferring to have it from the horse's mouth, is an original and in¯uential synthesis, Box and Jenkins 1976. 11.1

Time Series and Noise

If you have been reading straight through, you will by now have a fairly clear idea of what a time series is. But a little redundancy on important matters never hurts, and time series is the most important idea in the book. As de®ned by McCleary and Hay (1980, 30): ``A time series is a set of ordered observations: Y1 ; Y2 ; Y3 ; . . . ; Ytÿ1 ; Yt ; Yt‡1 ; . . .'' The key idea is that the observations are ordered in time, and this temporal order is denoted by the ubiquitous subscript. Any observed time series is considered to be a single realization of an underlying stochastic process (chapters 9, 12, and 13 discuss stochastic processes more generally). By modeling the process that generated the temporal order, linear time series analysis contrasts mightily with the most common practice of cognitive

90

Chapter 11

science research, which is to collapse observations across time, disregarding the order in which they were made. The basic stochastic process of linear time series analysis, and the only one we will consider here, is the ``white noise'' process (see also chap. 14). If we had a time series made up only of successive samples from a white noise process, a simple model of it would look like this: Yt ˆ b0 ‡ at ;

…11:1†

where b0 is a constant representing the level of the process, and at is a sample from a Gaussian (or normal) distribution with mean zero and constant variance sa2 . Because all such noise samples are assumed to be independent, at represents independent identically distributed, or i.i.d., noise. Also, the samples are assumed to be taken at equally spaced time points, although the actual time intervals could be anything from nanoseconds to many years, even unit intervals of ``system time.'' This is the null hypothesis of time series analysis. If a time series consists of, for example, a set of reaction times that can be modeled in this way, then the mean reaction time, which will equal b0 , will be a good representation of the situation and the temporal order of observations will be irrelevant: there is no temporal structure. All of the ¯uctuations we observe in such a reaction time series would be ``caused'' by the addition of independent samples of white noise, which have no temporal structure. Of course, as you will see in this and later chapters such is seldom the case. Either the noise is colored (see chap. 15), or there is some other, possibly nonlinear temporal structure in the series. Nonetheless, knowing for certain there is no temporal structure in the noise is better than making an unsupported assumption that there is none. 11.2

ARIMA … p; d; q†

The three processesÐautoregression (AR), integration (I), and moving average (MA)Ðthat combine to form the ``autoregressive integrated moving average'' of ARIMA are the ones used to model the generation of most linear time series. Each process is represented by a speci®c kind of linear model discussed below. A convenient convention to describe what kind of model best ®ts a particular time series is to write ``ARIMA …p; d; q†,'' where p, d, and q stand for the order of AR, I, and MA types of model, respectively. Whereas AR and MA models are rarely combined (see McCleary and Hay 1980 for a detailed discussion of the reason, and sec. 11.3 for an example of a combination), both types are commonly

Linear Time Series Analysis

91

combined with I models. After brie¯y describing the ®rst-order processes (higher orders are easily generalized) of each type of model, I will summarize the procedure used to identify which type applies to a given time series, using the autocorrelation and partial autocorrelation functions. The simplest departure from the null hypothesis model of equation 11.1 occurs when the underlying process adds, or integrates, successive samples of noise (called ``shocks'' in the parlance of linear time series analysis). The Brownian motion discussed in chapter 14 is a good example of such a random walk, in which a tiny particle starts out at some position in spacetime and each subsequent position is obtained by moving a random amount along each spatial axis. Thus the position of the particle at any time is represented by the algebraic sum of its separate movements along each spatial axis at all previous times. Such a process is modeled by Yt ˆ Ytÿ1 ‡ at ;

…11:2†

where each new value of the variable Yt is composed of the previous value plus a sample of white noise. One interesting aspect of the behavior of this model is that its time series show substantial drift, that is, what appear to be steady movements in one direction or another for long periods of time. Stock market prices are good examples of drifts. These drifts might be mistaken for trends, that is, situations in which real change is taking place in an underlying process over time. An example of a trend would be motor learning, for example, learning to ride a bicycle, in which the level of skill really is increasing over time, although large random ¯uctuations also occur. Equation 11.2 can be modi®ed to include an additive constant to account for such a trend, and a t-test can be done to determine whether it is necessary. If either a drift or a trend occurs in a time series, then it is not stationary and must be di¨erenced to remove the trend or drift before any other model is ®tted. To di¨erence a series, each observation is simply subtracted from the one immediately following, to create an ordered series of di¨erences: d1 ; d2 ; . . . ; dtÿ1 ; dt ; dt‡1 ; . . . This new time series can then be examined for further drifts or trends, possibly requiring more di¨erencing, or checked for applicability of AR or MA models. The number of times di¨erencing is done becomes the value of the d parameter in the ARIMA description. For example, a series rendered stationary by a single di¨erencing operation, and having no signi®cant AR or MA parameters, would be ARIMA (0, 1, 0). It is common for behavioral times series to require di¨erencing at least once because processes such as learning and fatigue, which would cause trends, are di½cult to avoid in cognitive science experiments, and processes that depend on each previous

92

Chapter 11

position in some state space and cause drifts, such as evidence accumulation, are also ubiquitous in such experiments. Although random shocks persist in a time series, the in¯uence of each one may decrease as time passes. This is probably the simplest case in which a stochastic process has a temporal structure worth modeling. In AR models, the decreasing in¯uence of random shocks is modeled by including a multiplicative parameter on previous observations. In the simplest case, this looks like Yt ˆ f1Ytÿ1 ‡ at ;

…11:3†

where ÿ1 < f1 < 0 or 0 < f1 < 1. Because f1 is a fraction, each shock has an exponentially decreasing e¨ect (f1n for a shock n time intervals back in the series) on the current value of the series. This behavior is called ``autoregression,'' hence the name of the model. A time series requiring a model like that of equation 11.3, would be an ARIMA (1, 0, 0) model; a series that had to be di¨erenced once before ®tting such a model would be an ARIMA (1, 1, 0) model. An ARIMA (2, 0, 0) would require inclusion of two previous observations, their respective multiplicative parameters, and so on for higher orders. Shocks persist inde®nitely in both I and AR models, albeit their in¯uence decreases exponentially in AR models. In MA models, a given shock persists only for a ®nite, usually short, number of observations. The simplest MA model is Yt ˆ at ÿ y1Ytÿ1 ;

…11:4†

where, again, ÿ1 < y1 < 0 or 0 < y1 < 1, and each shock persists for only one observation in the past. If shocks persist for two observations, we will need two parameters, one for each of the two previous observations, and so on for higher orders. The ®rst-order MA model is represented as ARIMA (0, 0, 1) and for a once-di¨erenced series ARIMA (0, 1, 1). We seldom see an ARIMA (1, 1, 1) model because AR and MA processes show parameter redundancy, that is, they are mirror images of each other mathematically (see McCleary and Hay 1980). Although we now know what linear, noise-driven models can be ®tted to time series data, we do not know how to accomplish the ®ttingÐthe task of model identi®cation, which relies on two statistics calculated on the time series at hand, the autocorrelation function (ACF) and the partial autocorrelation function (PACF). To statisticians, a correlation is a normalized covariance, namely, a covariance divided by a variance-like term, actually, the product of two standard deviations. The covariance (cov) of

Linear Time Series Analysis

93

a set of paired numbers, say …xi ; yi † representing two time series, is the average summed product of deviations of the separate members of each pair from their respective means, mx and my : P …xi ÿ mx †…yi ÿ my † ; covxy ˆ i N ÿ1 where i ˆ 1 to N. The variance (var) is the covariance of a variable with itself, for example, P …xi ÿ mx †…xi ÿ mx † ; varx ˆ i N ÿ1 where i ˆ 1 to N, and the standard deviation is the square root of the p variance, sdx ˆ varx . Thus correlationxy ˆ covxy =…sdx
sdy †. Notice that the product of standard deviations in the denominator is like a variance. Physicists usually refer to the summed product of the numbers in P each pair … i xi
yi †, possibly normalized by dividing by the number of pairs, as the correlation, because that is the core of the concept. Everything else is normalization required to make comparisons across contexts. Here I use the statisticians' de®nition, what the physicists call the ``dimensionless covariance'' because dividing by the product of standard deviations cancels out the physical (or other measurement-related) dimensions of the quantities involved. An autocorrelation, as you might have suspected (or already known) is a correlation of a time series with itself at a certain lag, which refers to the di¨erence in subscript values between each observation in the series and the observation at an earlier or later time step. For example, if we have a series 3, 5, 7, 5, 8, 2, 4 . . . then the autocorrelation for lag 1 is between the values 3, 5, 7, 5, 8, 2, 4 . . . and 5, 7, 5, 8, 2, 4 . . . , or alternatively the pairs of values (3, 5), (5, 7), (7, 5), (5, 8), (8, 2), (2, 4), (4, . . .) . . . The ACF consists of the set of autocorrelations at all possible lags, 0 through N ÿ 2 (because it is symmetrical, we usually consider only one direction). The autocorrelation at lag 0 is always 1 because it is the variance of the entire series divided by the same variance, and is unde®ned for lags greater than N ÿ 2 because at least two pairs are required for the variance to exist. Autocorrelations are estimated very well at small lags but are increasingly poorly estimated at higher lags as the number of pairs available becomes smaller. For short time series, none of the autocorrelations is estimated very well. The PACF is the set of partial autocorrelations at all lags, where the partial autocorrelation at a speci®c lag is the autocorrelation at that lag, with the autocorrelations at smaller lags ``partialed out'' (see again McCleary and Hay 1980).

94

Chapter 11

Conveniently, the various ARIMA models produce time series with unique ACFs and PACFs, so that model identi®cation can be accomplished by inspecting these functions for the time series of interest (see McCleary and Hay 1980 for the derivations). Of course, white noise alone produces a null ACF and PACF because, by de®nition, it is composed of independent samples from a Gaussian, or normal, distribution, whereas the ARIMA (0, d, 0) integrated process produces an ACF with signi®cant correlations at several lags, with the ACF gradually decreasing as lag increases. Di¨erencing such a process d times creates a null ACF and PACF. AR processes produce ACFs that decrease exponentially with lag and PACFs that have a signi®cant correlation at only the number of lags that represent the order of the process. MA processes create the opposite kind of picture, with ACFs that have signi®cant correlations only at the number of lags that represent the order of the MA process, and exponentially decreasing PACFs. Figure 11.1 shows examples of ACFs and PACFs for ARIMA (1, 0, 0) and ARIMA (0, 0, 1) processes, where either 0 < f1 < 1 or 0 < y1 < 1. The pictures for ÿ1 < f1 < 0 or ÿ1 < y1 < 0 are similar. The AR process with negative f1 has an ACF with correlations of exponentially decreasing absolute value but alternating signs at successive lags and the signi®cant PACF correlations are all negative. In the MA process with negative y1 , the ACF and PACF are both positive at all lags instead of negative as they are in ®gure 11.1. After using the ACF and PACF of a time series, in combination with these prototypical patterns, to accomplish model identi®cation, the parameters of the model can be estimated, the residuals (remaining variability after the model is ®tted) diagnosed (they should be white noise), and the model put into use (see again McCleary and Hay 1980). In sections 11.3 and 11.4, I illustrate how the linear time series approach can be applied to time series relevant to dynamical cognitive science. 11.3

ARIMA Model of Time Estimation

In the four experiments reported in Ward 1996, six human subjects produced 1,000-point time series consisting of the intervals between 1,001 successive presses of a computer keyboard spacebar. Endogenous time series were produced by having subjects attempt to press the spacebar at 1 sec, 3 sec, and ``natural'' or ``random'' intervals, without any timekeeping aids. In my ARIMA analysis of these time series, none of the time series was stationary (a chronic problem with psychological time series); all were therefore di¨erenced once to render them so. In three of the experi-

Linear Time Series Analysis 95

Figure 11.1 Autocorrelation and partial autocorrelation functions (ACFs and PACFs) for simple autoregressive integrated moving average (ARIMA) models. (After McCleary and Hay 1980.)

96

Chapter 11

ments, the ACFs and PACFs generally resembled the MA pattern shown in ®gure 11.1, so the best model was ARIMA (0, 1, 1) (including d ˆ 1 for the di¨erencing) in which random shocks persisted for only one observation, with average y1 ˆ 0:789 for 1 sec, 0.600 for 3 sec, and 0.605 for natural intervals. In one case, an additional MA parameter, and in another case a negative AR parameter, was required. All of the ®ts were very good, indicating that the MA model was appropriate for these data, and the ACFs and PACFs of the residuals of the ®ts for all series were uniformly null, indicating that the ®tted models were statistically adequate. Thus, when humans are attempting to reproduce a series of remembered eotemporal intervals, a random ¯uctuation introduced to one interval in¯uences only the next one, as we might expect from the fact that the end of one interval marks the beginning of the next one. The random instructions, on the other hand, produced some contrasting behavior: three subjects had null ACFs (after di¨erencing). Which means they actually were able to produce a series of intervals uncorrelated except for a trend or driftÐan ARIMA (0, 1, 0) model. It seems likely that there was indeed drift, again because the end of one interval marked the beginning of the next, thus we would expect drift from the sum of uncorrelated intervals. This performance was remarkable, given the di½culty most people have in producing a random sequence (see chap. 29). Perhaps the temporal interval context made it di½cult for these subjects to employ their usual ``randomness'' heuristics, and thus to produce their usual, correlated, series. The other three subjects in this experiment required ARIMA (0, 1, 2) models to describe their data, meaning that the ``random'' instruction induced them to create a correlation of their present interval with the intervals both one and two back in the sequence, a result to be expected from most people attempting to behave randomly. The ARIMA analysis clearly di¨erentiated between the random and the other instructions, although, for three of the subjects, it simply increased the persistence of the previous shocks. Importantly, even after all of this (interesting) linear temporal structure was removed from these time series, they still displayed evidence of nonlinear structure, perhaps chaos (see Ward 1996 and chap. 26). 11.4

Mixed Regression-ARIMA Model of Psychophysical Judgment

For a second, more complicated, example of the application of ARIMA analysis, consider the following time series made up of magnitude estimation stimuli, Sn (here pure tones varying in intensity), and re-

Linear Time Series Analysis

97

sponses, MEn (here numbers produced by human subjects that are supposed to be proportional to the perceived loudnesses of the tones): S1 ME1 S2 ME2 S3 ME3 S4 ME4 . . . Sn MEn . . . In a series of papers, I and several other researchers applied multiple linear regression to create empirical models of such time series. In this technique, the logarithm of each response is considered to be a joint function of the logs of current and previous stimuli and responses up to some lag. The logarithms of stimuli and responses are used to allow linear regression to be applied; the theoretical model on which the analysis is based resembles Stevens's psychophysical power law (e.g., eq. 1.3), with added terms for previous stimuli and responses. A good empirical model for most magnitude estimations can be expressed as follows: (Jesteadt et al 1977; Ward 1979): log MEn ˆ b0 ‡ b 1 log Sn ‡ b 2 log MEnÿ1 ÿ b 3 log Snÿ1 ‡ e;

…11:5†

where e stands for residual error. DeCarlo and Cross (1990) analyzed the same time series di¨erently, combining the regression technique with an ARIMA analysis. They argued that the dependency of the current response on the previous response should be seen in the context of ``correlated error,'' which is exactly the ARIMA assumption. Thus they ®rst regressed log Sn and log Snÿ1 on the current response, equation 11.5, with the ``b2 log MEnÿ1 '' term removed. The result of removing this term was to make the coe½cient of log Snÿ1 positive rather than negative as in equation 11.5, producing a very di¨erent regression equation. When DeCarlo and Cross then applied an ARIMA analysis to the residuals of the new regression equation, they found that an ARIMA (1, 0, 0) process ®t these residuals very well, implying that b 2 log MEnÿ1 in equation 11.5 could be replaced by f1 enÿ1 : log MEn ˆ b0 ‡ b 1 log Sn ‡ b 2 log Snÿ1 ‡ f1 enÿ1 ‡ e;

…11:6†

where f1 is the AR parameter. Notice that b2 is the coe½cient of log Snÿ1 in equation 11.6, and that it is now positive. Although both equations 11.5 and 11.6 ®t the data well, they have very di¨erent theoretical implications, which have still not been reconciled. DeCarlo and Cross argued that equation 11.6 is a better empirical model for such time series than is equation 11.5 because log Snÿ1 and log MEnÿ1 are highly correlated in the time series. (Indeed, in psychophysical scaling, most of the variance in the responses is produced by variations in the stimuli, meaning that the two variances are highly correlated.) Such correlations between regression variables cause serious problems for regressions, a problem that I and others dealt with by ®xing

98

Chapter 11

the order of entry of variables into the regression to be the assumed causal order. Although e¨ective, this approach is empirically inferior to dealing with only uncorrelated regression variables. DeCarlo and Cross's approach (1990) has this advantage, and also an e¨ective way, ARIMA, to deal with the structure in the residual variance. The question is whether to allow the theoretical stance to be dictated by the empirical model that is easiest to use. Because both theoretical approaches have been con®rmed by separate experiments and seem useful, the issue continues to be undecided. This controversy demonstrates that although ARIMA (and linear regression) can be extremely useful techniques, they are only tools, and limited ones at that, as we will see in chapters to come.

Chapter 12 Probability Theory and Stochastic Processes

Although gamblers have undoubtedly learned a great deal from probability theory, they clearly never learnÐor in any event never acceptÐthe main lesson probability theory can teach about gambling . . . ``If you keep on gambling, you will lose.'' ÐWarren Weaver, Lady Luck A model is a simpli®ed description of how something is supposed to work. We hope. A stochastic model says this something is working with the help of some random, or stochastic, component. It is a kind of cop-out. Really we don't know how this something works, exactly . . . so we make a stochastic model. ÐPriscilla E. Greenwood, lecture notes for mathematics 608

12.1

Dynamical Cognitive Science and Mathematics

As you might have gathered by now, I feel that mathematics should play a central role in dynamical cognitive science. Mathematics is the ``Queen of the Sciences,'' the preferred language of the most successful ones, such as physics, and the bane of the less successful aspirants. Or, to quote Feynman (1965, 40): ``Mathematics is not just another language. Mathematics is a language plus reasoning [Ðit] is a tool for reasoning'' (emphasis his). As in all science, algebra is ubiquitous in dynamical science; the role of calculus, especially that of di¨erential and di¨erence equations, has been extensively developed, as is only natural: the di¨erential calculus is about change, and so is the science of dynamics. On the other hand, the role of probability theory in dynamical science has been relatively neglected, probably because noise often has been considered a nuisance rather than an important process worthy of study (see chap. 14). This chapter is therefore devoted to revealing a few important aspects of a branch of probability theory that may have a profound bearing on dynamical cognitive science: nonlinear stochastic processes. The emphasis

100

Chapter 12

here is on the unfolding of these processes in time, and on what kinds of statements we can make about the general behavior patterns followed by such processes. In passing, some of the equations used in earlier chapters will be considered from a slightly more mathematical point of view. Chapter 13 will describe the success physics has had applying a similar approach; chapters 12±14 will describe what can be gained from adopting a stochastic modeling approach to dynamical cognitive science. 12.2

Stochastic Processes: A Random Walk to Ruin

To begin gently, I am going to describe a classical, simple stochastic process, one we have already met in a di¨erent guise in earlier chapters. It is formulated in a gambling context for interest, but has been applied in physics (Brownian motion) and other areas and lays the groundwork for later discussion (see Feller 1968). Consider a fair gambling game in which a gambler with a stake, s, competes against a gambling house with stake S, both in dollars. The gambler bets $1 on each play, wins that bet with probability p and loses it with probability q, where p ˆ q ˆ 0:5 (note p ‡ q ˆ 1). Think of a fair coin being tossed and the gambler wins if heads appears and loses if tails appears. The gambler plays until his or her stake is lostÐor the bank is broken. What is the probability that the gambler will be ruined? What is the probability distribution of the duration of the game (i.e., what are the various probabilities that the game will continue for various numbers of plays until one of the end states is reached)? What is the expected time until the game ends (i.e., until one or the other player is ruined)? This game is an example of a restricted random walk problem. In the random walk conceptualization of the problem, a particle moving about on a line in discrete jumps takes the place of the gambler. The particle starts at position s and jumps about the line until it reaches either of two other positions, the origin (0) or another position s ‡ S, called ``absorbing barriers'' (if there are no absorbing barriers, the walk is unrestricted). For the random walk, p and q are the probabilities of moving to the right (toward s ‡ S, or winning) or to the left (toward 0, or losing). When p ˆ q, the walk is symmetric. When p > q, the particle tends to drift toward the right; when p < q, the particle tends to drift toward the left. Now we will see why this section has the strange subtitle it does. Returning to the gambling game, let us analyze the fair version ( p ˆ q ˆ 0:5; see Weaver 1963 for a simpler and clearer but longer analysis). Imagine you are the gambler and you are at Monte Carlo, an establish-

Probability Theory and Stochastic Models

101

ment that has a very large stake, S. Because you begin with your stake, s, the possible outcomes after one play, each with probability 0.5, are that (1) you have s ‡ 1 and Monte Carlo has S ÿ 1 or (2) you have s ÿ 1 and Monte Carlo has S ‡ 1. With each successive play, your fortune grows or diminishes by random increments and decrements of $1, until sooner or later one of the end states is reached. To calculate the probability of your ruin in this game, PR , we set up a di¨erence equation: PR …s† ˆ 0:5
‰PR …s ‡ 1† ‡ PR …s ÿ 1†Š;

…12:1†

which follows from the two equiprobable outcomes of each play: that you will either win a dollar or lose a dollar. We also know some boundary conditions: that PR …0† ˆ 1, that is, ruin is certain if you have lost all of your money, and PR …s ‡ S† ˆ 0, that is, ruin cannot happen if you have won all of the money. We can guess that PR is linearly related to s (if we are wrong, the solution will not work): PR …s† ˆ a
s ‡ b. Inserting the values of PR …s† found at the boundary conditions into this linear equation, we have 1 ˆ a
0 ‡ b when you have lost all of your money, and 0 ˆ a
…s ‡ S† ‡ b when you have won all of Monte Carlo's money. From the ®rst of these two equations, we see that b ˆ 1; inserting this into the second gives a ˆ ÿ1=…s ‡ S†. Plugging these values for a and b into the original linear equation, we obtain the solution PR …s† ˆ

S ; s‡S

…12:2†

which yields some shocking predictions. For example, if you begin gambling with $1,000, and the house has $100,000,000 (a not unrealistic assumption for Monte Carlo), your probability of ruin is PR …$1;000† ˆ $100;000;000 o $100;001;000 ˆ 0:9999900001. Thus it is virtually certain that you will lose all of your money, even though the game is ``fair.'' If, however, the game is slightly favorable to the house, p < q (as is almost always the case), you will lose with even greater certainty because the random walk will ``drift'' toward ruin, as if the thin line you are walking on is tilted toward the left. If you are interested in gambling, or even in probability theory, you will probably want to know more about how long such ``fair'' games can be expected to last. The (®nite) expected duration of the random walk to ruin given a player's initial stake s, D…s†, satis®es a di¨erence equation similar to equation 12.1 when p ˆ q: D…s† ˆ 0:5
‰D…s ‡ 1† ‡ D…s ÿ 1†Š ‡ 1;

…12:3†

with boundary conditions D…0† ˆ 0 and D…s ‡ S† ˆ 0. The solution to

102

Chapter 12

equation 12.3 is D…s† ˆ s
S. Thus, in a friendly game where each player begins with $100 and they play until one or the other wins all of the money ($200), the expected duration of the game is 100  100 ˆ 10;000 plays. If each play took 30 seconds (to ¯ip the coin and exchange money), the game would last 30  10;000 ˆ 300;000 seconds or 83.33 hours. Almost four days of gambling entertainment for a mere $100, which I will wager is much longer than you would have predicted in the absence of such an analysis. For our wealthier player at one play per second at Monte Carlo, the game would last on average 1;000  100;000;000 ˆ 100;000;000;000 seconds or about 3,171 years. Of course this picture is changed if p and q are not equal. The di¨erence equation and its solution are more complicated because they must contain p and q. The game is shorter when p < q than when p ˆ q, and longer when p > q. Finally, the probability that a game will last any particular number of plays can be derived using a similar di¨erence equation approach (see Feller 1968). The random walk toward ruin I have just described can be expressed as a linear stochastic process in terms of the position of the particle (you) on the line at some (discrete) time (equivalent to the current amount of money in your pocket at Monte Carlo) as a function of where it was at the previous moment: Xn‡1 ˆ Xn ‡ en‡1 ;

…12:4†

where en‡1 ˆ 1 with probability p, and en‡1 ˆ ÿ1 with probability q. (This form emphasizes that the process is unfolding in time.) The process described by equation 12.4 is actually an example of the autoregressive (AR) process, which has the more general form: Xn‡1 ˆ y
Xn ‡ en‡1 ;

…12:5†

where y is a real number, and en‡1 is a random variable that can have any probability distribution whatever. Notice that this is similar to the model described by equation 11.3, an ARIMA (1, 0, 0) or AR(1) model, except that there ÿ1 < y < 0 or 0 < y < 1. This is actually an in®nite family of models or linear stochastic processes, each di¨erent value of y yielding a di¨erent model or process. Stochastic processes can be written even more generally as Xn‡1 ˆ f …Xn ; en‡1 †;

…12:6†

where n is a discrete index (time or space; may be continuous), the values of en‡1 are random, and the function f , called the ``parameter'' of the model, has some relevant properties, such as continuity, boundedness, and linearity. If the function f is nonlinear, then we have a nonlinear

Probability Theory and Stochastic Models

103

stochastic process, as in the following equation: Xn‡1 ˆ y
Xnb ‡ en‡1 ;

…12:7†

where b is also a real number. Because equation 12.7 is a nonlinear function, its parameter, (y, b), is similar to those of the logistic and other chaotic equations. In general, linear stochastic models are easier to analyze and are better behaved than nonlinear stochastic models, as are their deterministic counterparts. Nonlinear models can exhibit very bizarre, but interesting, behavior as the possibly already exotic behavior of the nonlinear deterministic part (called the deterministic ``skeleton'') is kicked around by the stochastic part (see sec. 24.3). Mathematicians like to describe stochastic processes in terms of the probability distributions that characterize them; these descriptions are called the ``laws'' of the processes (see Grimmett and Stirzacker 1992 for more on the mathematics of stochastic processes). In other words, they describe a stochastic process by specifying the probabilities with which the various possible states of the system, Xn , can occur. These probabilities are a¨ected both by the deterministic part and by the stochastic part of such a process. If their probability distributions do not change depending on where in the state space the processes are operating, then they are said to be ``stationary,'' and the laws ``stationary laws.'' For the random-walkto-ruin process, unfortunately, a stationary law does not exist, which means that the probabilities of the di¨erent possible states are continually evolving as the process continues. However, for the more general AR process of equation 12.5 with the en‡1 samples from a Gaussian distribution with mean zero and standard deviation s, and ÿ1 < y < 0 or 0 < y < 1, a stationary law does exist. Let us assume y ˆ 0:5. In this case, if we consider just the deterministic part of equation 12.5, Xn‡1 ˆ yXn , as n grows very large the process converges to 0 (e.g., if X0 ˆ 1, we have the sequence 1, 0.5, 0.25, 0.125, 0.0625, . . .). Whereas the stochastic part of equation 12.5 would keep the process jumping around, the deterministic part would cause it to drift always toward zero, so that the jumping around would be overlaid on this drift toward zero. The stationary law is just the distribution of the noise in this case (and whenever ÿ1 < y < 0 or 0 < y < 1), a Gaussian distribution as speci®ed. 12.3

Critical Points in Stochastic Models

If the AR process of equation 12.5 has a stationary law whenever ÿ1 < y < 0 or 0 < y < 1, what happens when y is outside these boundaries? Let us again consider only the deterministic part of the process,

104

Chapter 12

this time with y > 1, say y ˆ 2. For X0 ˆ 1, for example, we obtain the sequence 1; 2; 4; 8; 16; . . . ; and for X0 ˆ ÿ1, we get ÿ1; ÿ2; ÿ4; ÿ8; ÿ16; . . . ; clearly heading for ‡y and ÿy, respectively. When y < ÿ1, the sequences grow toward both ‡y and ÿy in sign-alternating jumps that grow larger and larger. For example, when y ˆ ÿ2 and X0 ˆ 1, we have 1; ÿ2; 4; ÿ8; 16; ÿ32; 64 . . . In any case, even though the stochastic part causes the process to jump around, unrelenting movement toward in®nity takes the process away from any particular neighborhood, so that it never settles down into a stationary law. What if y ˆ 1 or y ˆ ÿ1? Mathematicians call these cases ``critical points,'' at which stationarity is not even de®ned (so that the process could be considered not to have a stationary law at these points) because no probability measure is possible. In this case, no probability measure is possible because the variance of the process is not a constant, but rather increases as n increases (as it also does in the cases y > 1 or y < ÿ1). To see this, consider that X ei ˆ n var ei ! y as n ! y; var Xn ˆ var i

which makes the points y ˆ G1 interesting, but why are they ``critical''? They are critical because they divide the behavior of the AR process into two cases: behavior that possesses a stationary law (ÿ1 < y < 0 or 0 < y < 1) and behavior that does not (y > 1 or y < ÿ1). (Notice that if y ˆ 0, then this is just a simple white noise process, whose stationary law is that of the noise.) In general, the character of any stochastic process (eq. 12.6) depends on its parameter (the function f in our terminology), usually in a continuous manner. If there is a break, or discontinuity, in this dependency, the value of the parameter at that point is the critical point or critical value (``critical point'' refers more to the position of the process in the abstract parameter space at which the change in behavior occurs). Clearly, such critical points are important in the description of any stochastic or deterministic process, even a linear one. They are even more important in nonlinear processes because they tend to herald very exotic behavior. For example, in the deterministic logistic equation (eq. 1.4), the parameter value a ˆ 3:58 is a critical value. For a < 3:58, the equation behaves relatively tamely, jumping in a predictable sequence from one periodic point to another (although the number of periodic points depends on the value of a and can become quite large). However, for a > 3:58, chaos ensues, literally. For some values of a > 3:58, there is an in®nity of periodic points and the behavior of the equation is chaotic (see chap. 24). Notice that, in the case of the logistic equation, the critical

Probability Theory and Stochastic Models

105

point is not de®ned in terms of a stationary law, but rather in terms of a di¨erent aspect of the system behavior, the number of periodic points and the presence of chaos. Finally, let us consider the nonlinear stochastic process of equation 12.7. Does this process have any critical points? Here we must think in terms of a two-dimensional parameter space. Notice that the two parts of the parameter can work in opposition or can reinforce each other. Thus, in some parts of the parameter space, whenever y > 1 or y < ÿ1 and b > 1 or b < ÿ1, we have a situation similar to that of equation 12.5: yXnb will tend toward some in®nity and there will be no stationary law. When both ÿ1 < y < 1 and ÿ1 < b < 1 are true, on the other hand, there will be a stationary law, because yXnb will tend toward either zero or some other ®xed point or points under these conditions. If y > 1 or y < ÿ1 and ÿ1 < b < 1, or vice versa, things get interesting. Whether there is a stationary law will depend on the actual values of y and b, on which one wins the competition. In general, the values of y will have to be more extreme than those of b because multiplication is less powerful than exponentiation. Clearly, however, the critical points of this model, though complicated and not easy to ®gure out, remain important to characterizing the model's behavior. 12.4

Ergodicity and the Markov Property

Two other mathematical concepts that bear on stochastic processes should be introduced here. The ®rst, ergodicity, has to do with stationary

Figure 12.1 Transition matrix for fair coin toss game with equal stakes of $5.

106

Chapter 12

laws; the second, the Markov property (®rst encountered in chap. 2), with sequential behavior. Mathematicians call a stochastic process ``ergodic'' if it converges to a unique stationary law, no matter where in the state space it begins, that is, no matter what the initial probability distribution over the various states of the system is. The AR(1) process discussed in section 12.3 is ergodic because no matter what the initial position of the particle, the drift of the deterministic part toward zero would assure that eventually the stationary distribution would be reached. Of course, the farther from zero the initial position was, the longer on average it would take to reach the ®nal, stationary distribution. Mathematicians call a stochastic process a ``Markov process'' (after the mathematician Andrey Andreyevich Markov) if the next state of the system depends only on the present state, and not at all on any other states, previous or future (see Feller 1968). The random walk to ruin is a Markov process (or Markov chain in this case) in which each next state depends only on the present state (although the number of these might be very large). Consider the gambling situation in which two gamblers each have only $5 (let us say you are tossing a fair coin with a friend). Here there are only eleven possible states, and the probability of moving between them can be neatly summarized in a transition matrix (®gure 12.1) in which ``0'' and ``10'' are absorbing states (the transition matrix for player 2 is the same as ®gure 12.1, with ``player 2'' substituted for ``player 1'').

Chapter 13 Stochastic Models in Physics

Limited information about the dynamics of a system and uncontrollable in¯uences on the system often mean that it can only be considered as a statistical system. ÐJosef Honerkamp, Statistical Physics

13.1

The Master Equation

Physicists approach the modeling of intrinsically stochastic processes somewhat di¨erently from the way mathematicians do, as described in chapter 12 (see Plischke and Bergersen 1994; van Kampen 1985). Their approach, which is both very general and compatible with that of mathematicians, can also be of considerable use in dynamical cognitive science. Physicists usually begin with a description of a physical situation (naturally). The required description is a mesoscopic one, somewhere between the microscopic, in which noise consists of the ¯uctuations in position and momentum of each of a very large number of elementary particles, and the macroscopic, the level of classical Newtonian physics, in which noise is unimportant because there are so many particles that their individual ¯uctuations can be ignored. In a mesoscopic description, noise is important but it is treated globally, by describing its probability distribution. Usually, we think of physical systems as composed of particles in motion, such as electrons ¯owing in a wire. At the mesoscopic level, the states of this system could be the di¨erent numbers of electrons ¯owing in the conduction band at a given time, which means that the electrons are ``free'': they are not occupying any orbits around nuclei of atoms, thus can move along the wire. We are not interested in exactly which electrons are ¯owing, however, but only how many are ¯owing and how many are bound in orbits around nuclei. The number of electrons ¯owing is related to classical variables such as current, I, but the classical treatment considers only its mean, or characteristic value, as in Ohm's law: E ˆ IR, where E stands for electromotive force (voltage) and R for resistance.

108

Chapter 13

In the mesoscopic description, we assume that the system changes between states by stochastic transitions, that is, we assume that transitions between states occur probabilistically and on a short timescale relative to the timescale under study. In our electron example, there would be a certain probability that a transition between x and x ‡ 1 electrons in the conduction band would occur at any moment. Whether it actually occurred at a given moment would be the result of some stochastic process, probably inherently unavailable to our observation. Let q be the state of such a dynamical system. Figure 13.1A shows part of the temporal unfolding, or time evolution, of such a system. Let W …qnew j qold † be the rate of transitions between states qold and qnew . The type of model we are interested in is the one speci®ed by the entire set of rates W for all possible states qold and qnew . Our model description of the system is in terms of the probability of state q occurring at time t: P…q; t†. Because we are interested in how P…q; t† changes over time (the dynamical aspect), we write a master equation (e.g., Nordsieck, Lamb, and Uhlenbeck 1940) that describes these changes in the most general way possible: … qP…q; t† …13:1† ˆ ‰W …q j q 0 †P…q 0 ; t† ÿ W …q 0 j q†P…q; t†Š dq 0 ; qt where q and q 0 are any two states, and the inde®nite integration is over the entire set of possible states. Basically, what equation 13.1 says is that the rate of change of the probability that the system is in state q at any time t, P…q; t†, is equal to the weighted ``sum'' over all possible system states of the di¨erence between the likelihoods that (1) being in some other state with probability P…q 0 ; t†, the system will at that time make a transition into state q; and (2) already being in state q with probability P…q; t†, it will make a transition out to some other state. The weighting is by the transition rates that characterize each state. This partial di¨erential equation describes the time evolution of the probability distribution of system states, not that of the state of the system itself. This distinction is critical to understanding how to use this approach. What is of interest here is whether and, if so, when there is an equilibrium in the probability distribution, that is, whether a particular, stable, probability distribution comes to describe the state transitions of the system. This is analogous to the property of stationarity in the mathematical theory of stochastic processes (chap. 12). Notice that the master equation also describes a Markov process because it is assumed that the system state at the next moment of time depends only on the state at the current moment of time (chap. 12). Figure 13.1B displays an example of the time evolution of a probability distribution toward equilibrium.

Stochastic Models in Physics

109

Figure 13.1 (A) Time evolution of a simple physical system; (B) time evolution of a probability distribution toward equilibrium.

110

Chapter 13

Physicists consider the master equation to be ``solved'' if they can ®nd the probability distribution P…q; t† that satis®es equation 13.1 for arbitrary p values of the probabilities at time t ˆ 0. For example, P…q; t† ˆ …1= 2p† exp…ÿq 2 =2† indicates that the ®nal, stationary probability distribution is Gaussian, or normal. Although it is not always easy to solve the master equation, general methods have been worked out and many special cases have been studied (e.g., van Kampen 1976, 1985). The general method consists of substituting equations for the W rates into equation 13.1 and applying various boundary conditions and other facts. In the more complicated cases, computer simulations are required to explore the evolution of the probability distribution, especially for some of the stochastic processes of interest to cognitive scientists, such as multiplicative noise, self-similar processes, and fractal processes. Finally, if we consider only the general location of the probability distribution and ignore its spread (see ®g. 13.1B), the mesoscopic description is consistent with the macroscopic description in the form of classical-type laws (although these may be nonlinear). 13.2

Quantum Physics

The approach to statistical physics brie¯y outlined in section 13.1 is concerned with the evolution of the probability distribution describing the mesoscopic statistical ¯uctuations in a dynamical system. Quantum physics is concerned with the evolution of a related but quite di¨erent aspect of a particular dynamical system: the probability amplitude for a quantum system to assume a particular state. Quantum theory addresses a microscopic scale of energy, space, and time, and on that scale energy (at least) comes in little chunks (called quanta) when we measure it. To describe the evolution of a quantum system over time, however, we need to use a di¨erent concept, that of the wave function, represented by a lowercase psi with a subscript, cn . The ®rst postulate of quantum theory is that the state of a quantum system is described completely by the wave function, where the subscript n stands for quantum numbers, one for each coordinate of the space in which the wave function is de®ned (see Baggot 1992). The wave functions represent probability amplitudes for the quantum system to assume various possible states, such as for an electron to occupy particular regions of space or to move at particular speeds. The probability (or probability density for continuous, in®nite systems) is given by the square of the wave function, or, becausepthe  probability amplitude is a complex number (e.g., a ‡ ib, where i ˆ ÿ1), by jc 2 j ˆ

Stochastic Models in Physics

111

c  c, where c  is the complex conjugate of c (e.g., a ÿ ib). Although the theory can be quite complicated, for present purposes we do not need to go into these complications. In this section, we will simply consider an alternative way of thinking about how a system unfolds in time that has proven to be very useful in physics. In fact, it is the foundation of quantum electrodynamics (see Feynman 1985), one of the two most successful theories in physics, the other being Einstein's general theory of relativity. Another important postulate of quantum theory is that observable quantities (the quantities we measure in experiments) are represented by mathematical operators. Although it is easiest to think of the operators of normal arithmetic, such as addition or multiplication, the operators in quantum theory can be much more complex. Indeed, the quantum theory idea is that measurement consists of applying an operator to the wave function, where the operator is essentially a description of the experimental system in all of its aspects, including especially the way in which measurements are made. One of the most famous equations of physics, the SchroÈdinger wave equation, ^ ˆ Ec; Hc

…13:2†

when written this way, uses what is called the ``Hamiltonian operator,'' ^ Equation 13.2 mathematically describes the measurement of a system's H. energy, using two separate operators: the kinetic energy operator and the potential energy operator. Whereas the kinetic energy operator involves di¨erentiation, the potential energy operator usually involves only multiplication. In equation 13.2 (technically, an eigenvalue equation; see Baggot 1992), the Hamiltonian operator operates on a wave function to yield that same function times the system's total energy, E. The version of the SchroÈdinger equation in equation 13.2 is time independent: it does not describe the time evolution of a quantum system, but rather the system at some ®xed time, even though it contains a di¨erential operator for the kinetic energy (energy of motion). To introduce time into the SchroÈdinger equation, we need to change our notation slightly. Physicists use a notation invented by Dirac (1958), representing the state of a quantum system by putting special brackets around the wave function symbol: jcn i, and by calling it ``state vector.'' Expressing this state vector as a combination of more primitive vectors that de®ne a special kind of phase space called ``Hilbert space,'' lets us use some powerful mathematical ideas. Thus expressed, state vectors are represented by boldface capital psi without a subscript: jCi. We can now write the time-dependent form of the SchroÈdinger equation:

112

Chapter 13

q ^ jCi ˆ HjCi; …13:3† qt p where i ˆ ÿ1 and h ˆ h=2p, h being Planck's constant. This equation states that the rate of change of the system's state vector over time depends on the Hamiltonian operator's action on that state vector. Integrating equation 13.3 (see Baggot 1992) yields ih

^ 0 i; jCi ˆ UjC

…13:4†

where ^ U^ ˆ eÿiHt=h ;

…13:5†

jC0 i is the state of the system at some initial time t ˆ 0, and jCi is its state at some later time. The complicated exponential term represented by equation 13.5 is called the ``time evolution operator.'' With it, equation 13.4 describes the time evolution of a quantum system (how it unfolds in time) as continuous and deterministic in the probability amplitudes of all the di¨erent possible states of the system in a linear superposition (see again Baggott 1992). But then where do the probabilities come from? Most physicists think about them in this way: the ``act'' of experimental observation ``collapses'' or ``reduces'' the wave function, producing a single ``real'' quantum state from the set of evolving, superposed (and possibly partly imaginary) ``possible'' states with a probability equal to the square of the amplitude for the system to assume that state. That is, although we can only observe states of a quantum system with various probabilities, the system evolves between our observations deterministically, not probabilistically. The process by which observation creates real quantum states by reducing the wave function is a fundamental mystery, often called the ``quantum measurement problem.'' In one popular interpretation of the theory, the wave function only represents our knowledge of reality, and has no reality in itself. In another, the observation actually creates the reality. Whichever interpretation we accept (and there are still others; see Baggott 1992 or Penrose 1990), we are clearly dealing with a way of understanding a dynamical system very di¨erent from the way in which dynamical stochastic systems are understood. Some suggestions about dynamical cognitive systems are reminiscent of this approach, for example, as Neuringer and Voss (1993) see it, chaos theory both explains the unpredictability of human behavior and provides it with a deterministic foundation (see chap. 30). Thus the formulations of quantum physics could prove to be more generally useful, especially in dynamical cognitive science.

Stochastic Models in Physics

13.3

113

Complementarity Redux

Let us return to the time-independent version of the SchroÈdinger equation (eq. 13.2). The E in that equation is called the ``eigenvalue,'' and the wave function an ``eigenfunction,'' of the Hamiltonian energy operator. The requirement that the eigenvalue be real, that is, observable, whereas the eigenfunction and the operator can be complex (and have imaginary components), puts a considerable constraint on the mathematical form of the operator. Only a certain kind of operator, one with real eigenvalues for the indicated wave function, can be considered. Even with this restriction, however, when we want to measure the value (actually its expectation value, a kind of mean, called an ``observable'') yielded by applying an operator in the course of an experiment, there is no inherent limit to the precision with which that value can be measured. For example, we could, in principle, measure the energy of a particle emitted by a radioactively decaying atom as precisely as we wished. Suppose, however, we wish to measure a second property of our decaying atom, say the time at which that particle was emitted. To do so, the mathematics requires that the two operators to be applied (one for the energy and one for the time) commute. That is, it must not matter in which order the two operators are applied to the wave function that describes the decaying atomÐeither way, the resultant eigenvalues, the observables, must be the same. This is actually true neither for energy and time operators nor for position and momentum operators, which do not commute. Indeed, it does not seem to be true for any pair of operators applied to the same quantum system, which means that, in general, we can measure only one property of a quantum system (determine the expectation value of its operator) to arbitrarily high precision. The value yielded by application of any other operator will be increasingly uncertain as we more precisely determine the value yielded by the ®rst. Such pairs of observables are said to be ``complementary.'' A more direct way to think about complementary observables is in terms of what the operators represent, an observer's real impact upon a quantum system. For example, to observe the position and momentum of an electron, we might throw photons at it. When the electron scatters a photon so that it enters our eye or some other photon-detecting device, we can determine its position (at the moment it scattered the photon). The higher the energy of the photon we throw at the electron, the more precisely we can determine where the electron was. On the other hand, the higher the photon energy, the bigger the ``jolt'' it gives the electron, changing its momentum in an unpredictable way. If we use lower-energy

114

Chapter 13

photons, we will be more certain about the momentum of the electron because the random jolts will be smaller, but less certain about its position. Heisenberg (1927) expressed this as an ``uncertainty principle,'' based on the idea that measurement of the two quantities requires two di¨erent kinds of measurement device and that the use of the one interferes with the use of the other. Mathematically, this is expressed as two operators that do not commute. Neils Bohr had a very di¨erent interpretation of Heisenberg's uncertainty principle, which he made into a philosophical principle (see Baggott 1992 for an excellent discussion). Bohr felt that the uncertainty principle, and complementary observables in general, represent a limit on what we can know. For Bohr, it was the very concepts of the complementary observables that were interfering with one another; they represented two complementary, not con¯icting, ways of viewing the quantum system and the world. In Ward 1990 and 1992, I explored the notion of complementarity more fully for psychophysics; Baird (1997) has taken it even further, proposing complementary psychophysical theories. In chapters 6±10, we applied Weinberg's principle of complementarity to yield complementary approaches to understanding cognitive systems, a theme we will pursue in later chapters.

Chapter 14 Noise

The earliest noise phenomena discovered were thermal noise [by Einstein in 1906] and shot noise [by Schottky in 1918]. In our opinion, these two noise processes are still the prototypes of all observed characteristic noise phenomena in and out of thermal equilibrium. ÐKaroline M. van Vliet, ``Classi®cation of Noise Phenomena''

14.1

What Is Noise?

Experimental data in cognitive science are noisier than those in physics, but not as noisy as those in social or clinical psychology. By this I mean that much of the variance in the dependent variables remains unexplained by the independent (or manipulated) variables. For example, consider an experiment (described in greater detail in chap. 16) where subjects responded to the onset of a target stimulus with a key press. About 60% of the variance in the reaction times of individual subjects was accounted for by the fact that there were either one, two, or four response alternatives in di¨erent conditions. This can be expressed as a linear regression equation: ^
NR ‡ b^ ‡ e; RT ˆ m

…14:1†

where RT is median reaction time, NR is the number of response alter^ and b^ are constants to be ®tted, and e is ``error'' or natives available, m ``noise.'' This noise represents the 40% of variance in individual reaction times that remained ``unexplained.'' Such noise is responsible for the ``error bars'' around the points in a typical plot of mean data, where the bars represent the sample standard deviation or standard error, usually assumed to be estimates of parameters of an ``error distribution'' sampled from on each trial. As chapters 11±13 describe, these samples of noise are assumed to cause ¯uctuations in behavior over time. How could noise do this? And what is noise anyway? Why can it not be gotten rid of? Does it

116

Chapter 14

have any structure at all, particularly temporal structure? In what follows in this and the next few chapters, I address these and other questions in an attempt to make the study of noise a respectable pursuit in dynamical cognitive science. The error, or noise, in an experiment such as the one I just described is generally considered to have two sources. The ®rst and assumed source of most noise in cognitive science and any other behavioral science is ignorance of the details of the uncontrolled variation of independent variables that are causally related to the dependent variables. When there is some knowledge of uncontrolled variation of independent variables, the noise caused by them in the dependent variables can be ``partialed out'' statistically, thus providing an ``explanation'' for some more of the response variance. Even if we knew everything about all of the relevant variables in an experiment, however, there would still be an irreducible minimum of ``essential noise,'' whose sourceÐthe second sourceÐis the unknowable. The universe is a quantum system, as are all objects and events therein. Quantum systems necessarily evolve outside of our awareness, that is, we cannot know the details of their evolution (see, for example, Baggott 1992). A statistical prediction of the state of a quantum system is the best that can be done (at least in the orthodox interpretation of quantum mechanics; see chap. 13). Moreover, measurement entails uncertainty, and the Heisenberg uncertainty principle (see again Baggott 1992) describes the way in which even measurement at the quantum level involves uncertainty. In this case, and in many at the classical level as well, measured variables trade o¨: more certainty in the value of one entails less certainty (more error or noise) in the value of another. Ultimately, all noise is quantum noise because the precision in the measurement of any set of variables is limited by quantum noise. Measurements usually also contain much ignorance noise because all of the relevant variables are seldom controlled, even in precise physical measurement. Moreover, much of what is called ``excess noise'' in physics, that is, noise in excess of thermal or shot noise, is ``colored noise'' (see chap. 15), which is both exotic and generic in dynamical systems. Thus a discussion of physical noise and its sources and types will reveal valuable information about how to cope with and understand the behavioral and physiological noise that characterizes the results of experiments in cognitive science. One very useful distinction is that between external and internal noise (e.g., van Kampen 1987). External noise is caused by a force outside of a system acting on that system. A good example is the turbulence of the earth's atmosphere acting on the light rays from distant stars, causing

Noise

117

them to twinkle. This kind of noise has two properties: (1) it does not a¨ect the actual behavior of the system of interest (the amount of light generated by a distant star is not a¨ected by the atmosphere's turbulence); and (2) it is connected to that system by some coupling mechanism and, in principle, can be switched o¨ (or by passed, as when the Hubble telescope in earth orbit views the star from outside the earth's atmosphere). In contrast to external noise, internal noise is an integral part of the system of interest, and participates completely in its temporal unfolding. ``It originates from the fact that matter consists of particles, or, as in photons, from the fact that energy is quantized.'' (van Kampen 1987, 4) A good physical example is the current noise described in chapter 19: the electrical current in a wire with a constant voltage source (say a battery) ¯uctuates randomly. According to van Kampen (1987), such internal noise often arises from the cooperative e¨ect of many system elements (e.g., electrons in current noise) rather than from ¯uctuations in the behavior of individual elements. One useful way to describe all external noise, and also internal noise in a system whose temporal unfolding is described by a linear equation, is with what is called a ``Langevin equation,'' after one of the early physicists who studied Brownian motion, the random motion of particles in a ¯uid caused by thermal noise (such as dust motes dancing in the air of a sunlit roomÐa type of internal noise). An example of such an equation is the one Langevin applied to Brownian motion: dv ˆ ÿbv ‡ l…t†; dt

…14:2†

where dv=dt is the momentary rate of change in the velocity, v, of a particle embedded in a ¯uid, b is the coe½cient of friction of the ¯uid, and l…t† is the force exerted by the other molecules in the ¯uid on a given particle. Langevin's insight was to describe l…t† stochastically, that is, to describe the in¯uences of the other particles on the motion of a given particle as a probability distribution, in this case described by only its mean (assumed to be zero) and its variance, which depends in turn on a few other parameters. The resulting equation successfully describes the ¯uctuations of Brownian motion in any ¯uid as long as the friction coef®cient is known. Thus a typical Langevin equation contains a deterministic term, which describes the temporal unfolding of the system, and a stochastic term, which describes the stochastic (or noisy) in¯uences of outside forces on that unfolding. In the case of our dust mote, the (linear) deterministic term describes the e¨ect of friction (external noise) on the

118

Chapter 14

particle's motion, whereas the stochastic term describes the e¨ect of other particles in the system (internal noise). Friction slows the particle, whereas other particles in the system cause no net change in the particle's motion, although, from moment to moment, they can either speed or slow the particle, sometimes by a large amount. Notice the resemblance between equations 14.1 and 14.2. Linear regression equations are a type of Langevin equation written in the form of a di¨erence equation (although eq. 14.1 is not a dynamical equation because time does not enter into that equation). Unfortunately, a Langevin approach does not su½ce for all types of internal noise. If the temporal unfolding, or time evolution, of a system is described by a nonlinear equation, then an attempt to model the internal noise using a Langevin approach leads to contradictions (van Kampen 1987). In this case, we have to use what is called a ``mesoscopic'' description of the system. ``It di¨ers from the microscopic level in that it is coarse-grained. . . . It di¨ers from the macroscopic level . . . in that the state of the system is not speci®ed by certain values of these variables, but by a probability distribution of them. The evolution is expressed by the time dependence of this distribution function'' (van Kampen 1987, 7). This approach, using the master equation, has already been described in some detail in chapter 13. Here it su½ces to point out that most of the interesting systems we will study in this book are nonlinear ones with internal noise, so that, in principle, the master equation approach applies to them, and especially to systems generating ``colored noise'' (see chap. 15). Another useful typology of noise was suggested by van Vliet (1981), who argued that internal noise could be classi®ed according to four principal distinctions: (1) thermal versus nonthermal; (2) microscopic versus mesoscopic; (3) lumped (discrete) versus distributed (continuous); and (4) classical versus quantum (all noise is ultimately quantum noise, but sometimes a classical description su½ces). According to van Kampen (1987), these four distinctions correspond roughly to which microscopic constant of nature determines the noise amplitude. Regarding the ®rst distinction, the strength of thermal noise always involves Boltzman's constant, k, multiplied by the temperature, T. For example, the variance of the stochastic variable l…t† in equation 14.2 is proportional to kT. The Brownian motion described by equation 14.2 is a prototypical example of thermal noise. van Vliet (1981) also suggested a superdistinction between characteristic and noncharacteristic noise. Characteristic noise is reducible to elementary events with time constants and can be described by a Langevin,

Noise

119

master, or other elementary equation. Noncharacteristic noise is not so reducible: there are no ``elementary'' events involved, and the ¯uctuations arise as a cooperative phenomenon. An example of noncharacteristic noise, according to van Vliet, is a particular variety of ``colored noise'' called ``1/ f noise'' (discussed in considerable detail in chaps. 16±19). The important point to keep in mind about noncharacteristic noise is that it cannot be described by conventional means. And the important overall concept to remember is that noise, though it might seem undesirable, is ultimately a fundamental aspect of the very phenomena we are studying; as such, it must be described with the same care as any other fundamental aspect. 14.2

Probabilistic Description of Noise

Because careful descriptions of noise often involve probability distributions and their properties, I will here brie¯y review two of the most commonly used probability distributions, plotted in ®gures 14.1A and 14.1B. (If you ®nd this material overly di½cult, you might consult the ®rst few chapters of a good textbook on probability theory, such as Feller 1968, for a gentler introduction to these concepts.) Whereas sample spaces with a ®nite number of outcomes assign probabilities directly to each outcome, continuous probability distributions are characterized by probability density functions that assign a probability density, rather than a probability, to each of an in®nite number of possible outcomes, or sample points. Probabilities in such in®nite sample spaces are integrals of the probability density function over some (still in®nite) set of sample points. The integral of the density over all sample points must equal one. The distribution function is the integral of the density function regarded as a function of its upper limit of integration, and assigns a probability to each of a series of sets of sample points. Often the expected value and the variance of a probability distribution su½ce to characterize it, but sometimes other moments must be used. When noise is characterized as a series of independent samples from a set of identical probability distributions, we say that it is ``independent, identically distributed'' (i.i.d.) noise. The simplest probability distribution is the uniform distribution, which has the same probability density, p…x† ˆ 1=…b ÿ a†;

…14:3†

for every possible sample point, x, where a and b are the lower and upper bounds of the real interval over which the probability density function is

120

Chapter 14

Figure 14.1 (A) The uniform probability distribution; (B) the standard Gaussian probability distribution.

Noise

121

de®ned. Equation 14.3, the uniform probability density function, is plotted in ®gure 14.1A. The most common uniform distribution is the one de®ned over the unit interval [0,1], so that p…x† ˆ 1 in the interval [0,1] and 0 elsewhere, and …1 …1 p…x† dx ˆ 1 dx ˆ 1: 0

0

The expected value of this distribution is 0.5 and the variance is 1/12. One very useful property of the uniform distribution is that if the interval ‰b ÿ aŠ is partitioned into n identical subintervals, the probability of a random sample from the distribution falling into a given subinterval is the same for each subinterval and equal to 1=n. The distribution function of the uniform distribution is a straight line whose slope is 1=…b ÿ a† and whose y-intercept is a. Thus the distribution function of the unit uniform distribution is a straight line whose slope is 1 and whose y-intercept is 0. Probably the most commonly used probability distribution is the Gaussian distribution (named after Carl Friedrich Gauss, the great mathematician who invented it) or ``normal'' distribution. The famous ``bell curve'' of the Gaussian describes so many aspects of the natural world it is small wonder that it also describes noise in so many dynamical systems. Its density function is 2 1 2 p…x† ˆ p
eÿ…xÿm† =…2s † ; s 2p

…14:4†

where the parameters m and s are the expected value and standard deviation, respectively. Its distribution function unfortunately cannot be evaluated analytically and must be approximated numerically. Tables of the distribution function of the standard normal distribution, which has m ˆ 0 and s ˆ 1, are ubiquitous, and approximation schemes abound. The standard normal probability density function is plotted in ®gure 14.1B. Notice that, since m ˆ 0 and s ˆ 1, the standard normal probability density function can be written 1 2 p…z† ˆ p
eÿz =2 ; 2p

…14:5†

where z ˆ …x ÿ m†=s. Thus the abscissa of the graph in ®gure 14.1B is in units of z. The Gaussian distribution has many wonderful properties, perhaps the most profound of which is that it is the limiting form of many other distributions. For example, the Poisson distribution, which characterizes radioactive decay, emission of photons by a light ®lament, and

122

Chapter 14

many other random processes, becomes Gaussian in the limit as its rate parameter approaches in®nity. Moreover, the distribution of sums or means of samples from any distribution, including the uniform distribution, becomes more like the Gaussian distribution as the number of samples increases. This creates both a popular and easy way to simulate the Gaussian distribution, namely, by summing samples from a uniform distribution, and a popular rationale for assuming Gaussian ¯uctuations in many situations. Any normal distribution can be transformed into the standard normal by applying the z-transform described in the context of equation 14.5, for example, when researchers simulate sampling from the standard normal distribution by summing samples from a uniform [0,1] distribution. If, say, they take a sum of twelve samples from the uniform xi are the samples distribution, that is, y ˆ x1 ‡ x2 ‡


‡ x12 , wherepthe  from the uniform, then m ˆ 0:5…12† ˆ 6 and s ˆ …12=12† ˆ 1, so that z ˆ … y ÿ m†=s ˆ … y ÿ 6†=1 ˆ y ÿ 6. The value of z represents a single sample from an approximately standard normal distribution. 14.3

Spectral Properties of Noise

Random noise can also be described by its power spectrum, although this was originally thought to be a meaningless exercise because random noise was supposed to have no periodic structure (see, for example, Bell 1960). A power spectrum is based on the Fourier transform, which expresses the relationship between time domain and frequency domain descriptions of any physical process. The Fourier transform of a function of time, h…t†, is written … ‡y h…t†
e 2pift dt; …14:6† H… f † ˆ ÿy

where H… f † is a function of frequency, f , measured in Hertz (Hz). The square of the Fourier transform, H… f † 2 , is called the ``power spectral density,'' and expresses the amount of temporal ¯uctuation that can be described by a combination of sine and cosine waves at a particular frequency f . The power spectrum of a function of time is a plot of the squared Fourier transform versus frequency, usually in log-log coordinates for convenience (i.e., log H… f † 2 versus log f ). The power spectrum of real, discretely sampled time series is usually estimated by a method called the ``fast Fourier transform'' (FFT), which gives the power spectral density in each of several small bands of frequencies into which

Noise

Figure 14.2 (A) Time series of Gaussian noise; (B) power spectrum of Gaussian noise.

123

124

Chapter 14

the relevant frequency range is divided, usually referred to by indicating the center frequency of the band (see, for example, Press et al. 1992). For example, consider a series of 1,024 samples from a Gaussian probability distribution. Although this is a relatively short time series, it is in the range of time series lengths to be expected in cognitive science experiments, and it is su½cient to illustrate the expected power spectrum of Gaussian noise for such time series. I generated such a time series (using a computer of course) and did an FFT to obtain the power spectrum (averaging the higher frequency bands into larger bins containing several adjacent bands as suggested by Mandelbrot 1998). The time series is displayed in ®gure 14.2A and its power spectrum in ®gure 14.2B. The power spectrum is relatively ¯at (slope ˆ ÿ0:004) over the range of frequencies involved, which means that all the frequencies contributed roughly equally to the ¯uctuations of the time series. The ideal power spectrum for such ``white'' noise has slope equal to zero. Of course, because any real system can have only a ®nite amount of power (the area under the power spectrum curve), this and any other spectrum must eventually have both a high- and a low-frequency cuto¨, where the spectral power density becomes zero. Descriptions of noise in terms of its power spectrum have become commonplace, and we will be using such descriptions routinely in the next several chapters, as we continue our search for temporal structure in noise.

Chapter 15 Colored Noise

When the ticks in my Geiger counter exhibit 1=f noise I must conclude that the nuclei in¯uence each other's decays, or that they remember the supernova explosion in which they were born, or that I have made a mistake. ÐN. G. van Kampen, ``Some Theoretical Aspects of Noise'' If you have not found the 1=f spectrum, it is because you have not waited long enough. ÐStefan Machlup, ``Earthquakes, Thunderstorms, and Other 1=f Noises''

15.1

The Ubiquity of Colored Noise

What do a Bach concerto, alpha waves, heartbeats, earthquakes, thunderstorms, electrical current, reaction times in cognitive science experiments, mistakes, and chaos all have in common? The answer is that time series of ¯uctuations characteristic of these phenomena all exhibit a 1=f power spectrum. Here I explore the meaning of this answer; in later chapters I discuss possible reasons why it is true. My opening question, far from being merely rhetorical, is both substantial and profound and leads to a second such question. Why is it that phenomena so diverse all share the same power spectrum of ¯uctuations, yet other, seemingly similar phenomena, such as Brownian motion and white noise (see chap. 14), do not? As noted in chapter 14, white noise is a time series of ¯uctuations having roughly equal spectral density at all frequencies in a given range. A good example of white noise is a time series of samples from a Gaussian, or normal, distribution. Figure 14.2 showed that the power spectrum of such a time series is roughly ¯at (zero slope in log-log coordinates). What this means is that there are no correlations whatsoever between samples

126

Chapter 15

taken at one point in the time series and those taken at any other; the samples are independent and identically distributed (i.i.d.). Although a single strong periodicity in the time series will of course be displayed as a single peak in the power spectrum at the frequency of the periodicity, the correlations in the time series of ¯uctuations from a Bach concerto, alpha waves, heartbeats, and so forth occur at several timescales, from very long range to very short range, with no one preferred timescale. Using a computer and equations 16.2±16.5 I created a time series of such ¯uctuations, which is displayed in ®gure 15.1A, where you can see the long-term correlations in the time series as gentle undulations, overlaid by higher-frequency ¯uctuations. The power spectrum of this time series, from a fast Fourier transform (FFT) and in a log-log plot, is shown in ®gure 15.1B. The slope of the best-®tting straight line through this power spectrum, treated in log-log coordinates, is not zero, but roughly equal to ÿ1. Noise with this spectral ``shape'' is called ``pink noise'' because it is dominated by relatively low frequencies, but has some higher frequencies in itÐjust as light that appears pink to human eyes is dominated by lowfrequency (long-wavelength) photons, but has some higher-frequency (short-wavelength) photons in it. As the supposed signature of a very interesting type of dynamical system called a ``complex system'' (Waldrop 1992), pink noise has proved notoriously di½cult to explain (see chaps. 18 and 19). I will have more to say about complex systems, but for now it su½ces to point out that, wherever pink noise appears, it is reasonable to look for other exotic phenomena, even in some cases the one called ``chaos'' (see chap. 23). Another type of ``colored noise'' (noise whose power spectrum is not ¯at), and one that is easier to understand, is ``brown noise.'' Probably the most famous example of brown noise is Brownian motion, which can be represented by a running sum of a time series of samples from a probability distribution with mean zero. The sum represents the current position of a particle bu¨eted by other particles in a ¯uid (see chap. 14); it is a kind of random walk. By sampling from a Gaussian distribution and adding the successive samples to create a running sum, I created some hypothetical Brownian motion. The time series of positions (momentary values of the running sum) is shown in ®gure 15.2A, and its power spectrum, again from an FFT and in a log-log plot, in ®gure 15.2B. It is clear that low-frequency ¯uctuations are even more dominant in brown noise, to the point where the series stays in the same neighborhood for relatively long periods of time. The power spectrum re¯ects this dominance by low

Colored Noise

Figure 15.1 (A) Time series of pink noise; (B) power spectrum of pink noise.

127

128

Figure 15.2 (A) Time series of brown noise; (B) power spectrum of brown noise.

Chapter 15

Colored Noise

129

frequencies, since the slope is about ÿ2 in the log-log plot (this result can be derived theoretically; see, for example, van Vliet 1981). Some noise is even darker than brown noise; the slopes of the log-log plots of the power spectra of such ``black noise'' are more negative than ÿ2 (Schroeder 1991). One dramatic and important example is the ¯ooding of the Nile river, whose ¯uctuations have a power spectrum with a slope of ÿ2.8. Such a spectrum means that the minimum yearly level of the Nile (recorded from 622 to 1469 c.e. by the Egyptians) changes only very slowly over periods of many years, leading to both droughts and wet centuries and profoundly a¨ecting human welfare in the region. These represent only a few of the many examples of colored noise in nature. More will be discussed later in this chapter and in chapters 15 and 16. For now, remember that far from being an exotic rarity in our world, colored noise is so common, and occurs in such an interesting array of dynamical systems, including cognitive systems, that its presence demands exploration and explanation. 15.2

The Vicissitudes of the Exponent a

One convenient and economical way to describe the power spectra of colored noises is with an equation such as the following: S… f † ˆ kf ÿa ;

…15:1†

where S… f † is the power spectral density (say, from an FFT), f is temporal frequency (in Hertz), k is a constant determined by speci®cs of the system such as the mean and variance of the noise, and a is the crucial parameter that determines the ``color'' of the noise. When we take the logarithm of both sides of equation 15.1, as we did when plotting the power spectra in ®gures 14.2B, 15.1B and 15.2B, we obtain log S… f † ˆ log k ÿ a log f ;

…15:2†

a linear equation in the variables log S… f † and log f , with slope ÿa. Thus, for pink noise ÿa ˆ ÿ1 (®g. 15.1B), for brown noise ÿa ˆ ÿ2 (®g. 15.2B), and for black noise ÿa < ÿ2. White noise has ÿa ˆ 0, implying that log S… f † ˆ log k ˆ constant (®g. 14.2B). Because the minus sign is somewhat awkward to carry around, in what follows I am going to drop it in front of a in most cases, but you should remember that it is supposed to be there. Moreover, as is customary in the ®eld, I will from now on write ``1=f noise'' when a ˆ 1, omitting the exponent.

130

Chapter 15

It would be an ideal world if only a few, integer values of a occurred, like the values taken by the quantum numbers associated with atoms. And such precise and sparse values do occur in some physical theories. For example, as mentioned above, it can be shown theoretically that the power spectrum of Brownian motion has a ˆ 2 (and thus ÿa ˆ ÿ2). But the real world, unfortunately, is not that neat. Many di¨erent values of a appear in di¨erent systems, and it is not yet clear which factors are responsible for these variations. Some theories in physics seem to explain some of this variation. For example, Bell's explanation (1960) for pink noise that occurs in electrical current in a wire also explains why the exponent of frequency varies somewhat with the type of material of which the wire is made (but see van Kampen 1987 and chap. 19). There are, however, no accepted theories of pink noise in living systems or in cognitive systems; even physicists consider pink noise to be a mystery: a steady stream of papers in physics journals contain models their authors claim reveal why pink noise is so ubiquitous (e.g., Kaulakys 1999). One reason for nonideal exponents appearing in equation 15.1 (or for nonideal slopes in eq. 15.2) is that we cannot observe time series for an in®nite time. This is especially important for higher values of a, such as those greater than two (or ÿa < ÿ2), which describe black noise. Black noise could characterize seemingly catastrophic phenomena other than Nile ¯ooding, such as ¯uctuations in global temperature or in the numbers of living members of endangered species of animals (Schroeder 1991). The power spectra of such phenomena might be only asymptotically black, approaching, say, an exponent of a ˆ 3 according to the following equation (Schroeder 1991): S… f † ˆ

T 4f ; 1 ‡ …T 4 f 4 †

…15:3†

where T is the time period over which the phenomenon is observed. As T gets very large, the 1 in the denominator of equation 15.3 becomes irrelevant and the equation approaches S… f † ˆ 1=f 3 ˆ f ÿ3 . However, when T is small, the equation ¯uctuates wildly for di¨erent values of T, with the extreme ¯uctuations depending on just how long T is. If we wait long enough, anything is possible (as in the second epigraph above). If the observation period is long enough, we can be fooled by one of these extremes (e.g., several unusually hot summers) into thinking that the entire process has changed (permanent global warming) when it may be, in fact, only an unusually ``black'' part of the black noise time series. It is

Colored Noise

131

thus very di½cult to discriminate permanent shifts in a time series (nonstationarity) from expected extreme ¯uctuations in black noise in which the underlying generating process has not changed at all. A measure of ¯uctuations that is especially useful for black noises, where the distance between di¨erent states appears to grow larger with time, is called the ``rescaled range'' and is symbolized by R=S (Schroeder 1991). The range, R, is the maximum value minus the minimum value of the variable in a stationary time series over the time interval observed, whereas the scaling factor, S, is the sample standard deviation of the set of values in the time series over that same interval. The range is a ``®rst power'' measure of the dispersion of values in the time series, whereas the standard deviation is based on a ``second power'' measure of the dispersionÐit is the square root of the sample variance, where Var ˆ

nÿ1 X

…xi ÿ x† 2 =…n ÿ 1†:

i

For colored noise, the rescaled range is proportional to the time interval, Dt, raised to a power, H: R ˆ Dt H : S

…15:4†

The exponent H, called the ``Hurst exponent'' after its inventor H. E. Hurst, can be easily measured for a particular time series as the slope of a plot of log R=S versus log Dt, in the same way as is done for the power spectrum. Importantly, for any time series, a ˆ 2H ‡ 1;

…15:5†

so that the exponent of the power spectrum can be obtained from the Hurst exponent or vice versa. Thus Nile ¯ooding, with a ˆ 2:8, has a Hurst exponent of 0.9, brown noise has H ˆ 0:5, pink noise has H ˆ 0, and white noise has H ˆ ÿ0:5. The Hurst exponent is a simple measure of persistence, or memory, namely, how long we can expect a given ¯uctuation in a time series to be re¯ected in future values of the series. The more positive the Hurst exponent, the longer the persistence (memory). Another way to talk about this is in terms of relaxation time, which is the time it takes for a system to ``relax'' from the e¨ect of a given ¯uctuation into its equilibrium, or steady, state (which it may never do because there will be future ¯uctuations to relax from). Relaxation times will appear prominently in many discussions in the chapters to come.

132

Chapter 15

Figure 15.3 Biological pink noise from the human heart. (From Musha 1981.)

15.3

Colored Noise in Living Systems

To many people, the most intriguing aspect of colored noise, especially pink, or 1=f , noise, is its appearance in both nonliving and living systems. Ever since SchroÈdinger (1944/1967) wrote about the paradoxes entailed by consideration of living things as physical systems, a bridge between nonliving and living systems has been sought. Pink noise may be a clue to nature of the long-sought bridge (or it may be a red herring that will lead us astray). If the underlying dynamic processes giving rise to pink noise in nonliving systems can be understood, perhaps this understanding can be extended to living systems. The same reasoning can be extended, in turn, to cognitive dynamical systems, ultimately to consciousness and the mind. Perhaps pink noise can reveal something fundamental about mental processes. This idea is considered seriously in chapters 16 and 17. Two good examples of pink noise in living systems come from studies of muscles. Musha (1981) reported a study of ¯uctuations of the time intervals between successive R peaks (re¯ecting muscle contractions) of the electrocardiogram of the human heart. The power spectrum of an example time series plotted in log-log coordinates is shown in ®gure 15.3. As the ®gure makes clear, the power spectrum is just about 1=f (i.e., the log-log slope A ÿ1) over several decades of frequency. Figure 15.3 is a good example of power spectra from ``real'' time series. Notice it has a small peak at about 0.3 Hz, the frequency at which the subject was breathing, and clearly it in¯uenced the heart's contractions. Musha

Colored Noise

133

(1981) also reported an early study of the postural sway of the body on a platform, the slope of the power spectrum, plotted log-log, was ÿ1 for frequencies lower than 1 Hz, and slightly greater than that for higher frequencies. Musha suggested that the power spectrum is related to the mechanism of posture control, and more recent studies have con®rmed this (Lauk, et al. 1998). Biological pink noise also occurs both in the brain (chap. 17) and in the behavior of the whole humanÐto which we turn next.

This page intentionally left blank

Chapter 16 1=f Noise in Human Cognition

The [1=f music] was judged . . . to be much more pleasing than . . . (either white noise or 1=f 2 music). . . . Indeed, the sophistication of this ``1=f music'' . . . extends far beyond what one might expect from such a simple algorithm, suggesting that a ``1=f '' noise (perhaps that in nerve membranes?) may have an essential role in the creative process. ÐRichard F. Voss and John Clarke, `` `1=f Noise' in Music and Speech''

16.1

Music and Time Perception

The search for the relevance of 1=f noise to human cognition began with Voss and Clark 1975, quoted above. Although the paper was ignored for twenty years, it ®nally came to the attention of psychologists and has stimulated a recent burst of work in several laboratories. What Voss and Clarke did was to create several time series by digitizing the analog ¯uctuations of sound intensity and fundamental frequency (well below the ``audio'' frequencies of the sound waves themselves) in recorded music and voices played over the radio. Then they performed fast Fourier transforms (FFTs) on the time series to obtain the power spectra of the ¯uctuations (see chaps. 14 and 15). Their most famous result is shown in ®gure 16.1, the power spectrum for Bach's Brandenburg Concerto no. 1. Although two small peaks are apparent at about 2.5 Hz and 5.6 Hz, caused by the rhythmic structure of the piece, the spectrum below 1 Hz (``0'' on the log10 … f † axis) is approximately 1=f . Similar results were obtained for average time series of sound intensity ¯uctuations for Scott Joplin piano rags, classical music, rock music, and a news and talk station. The results for ¯uctuations of fundamental frequency were somewhat di¨erent, with only classical music exhibiting a broad-range 1=f power spectrum. The news and talk spectrum was the most deviant, with peaks representing the average length of an individual speech sound (0.1 sec)

136

Chapter 16

Figure 16.1 Power spectrum of amplitude changes in Bach's Brandenburg Concerto no. 1. (From Voss and Clarke 1975. Reprinted with permission from Nature. ( 1975, Macmillan Magazines, Limited.)

and the average time for which a particular person talked (100 sec). The meaning of speech is only weakly correlated with the structure of fundamental sound frequency changes, whereas that of music is usually carried directly by such changes, which comprise the melody (but see Boon and Decroly 1995). On the other hand, important meaning is carried by sound intensity ¯uctuations in both music and speech. And that meaning seems to be most esthetically pleasing when the intensity ¯uctuations are roughly 1=f . Independently of Voss and Clarke, Musha (1981) summarized several studies establishing a role for 1=f noise in biological systems, particularly in human perception. His own experiments established that the power spectra of some cartoons (based on spatial frequency ¯uctuations) are approximately 1=f , whereas those of some ``realistic'' paintings and photographs are 1=f 2 , and some impressionistic paintings have spectra somewhere between the two. Takakura et al. (1979) found that the painrelieving e¨ectiveness of transcutaneous electrical stimulation was doubled (from 35% to 70%) by synchronizing the electrical pulses with 1=f noise. Thus 1=f ¯uctuations are associated both with pleasure and with the absence of pain. In the ®rst of the more recent studies of 1=f noise in human behavior, Gilden, Thornton, and Mallon 1995, human subjects were asked to estimate various time intervals repeatedly (only one at a time) for many linked intervals; time series were constructed of the errors of the estimates

1=f Noise in Human Cognition

137

Figure 16.2 (A) Power spectra of ¯uctuations in time estimations; (B) power spectra from a model. (From Gilden, Thornton, and Mallon 1995. Reprinted with permission from Science. ( 1995, American Association for the Advancement of Science.)

(target time interval minus estimate for each estimate). When Gilden, Thornton, and Mallon performed FFTs on these time series, they found the average power spectra shown in ®gure 16.2A. The numbers labeling the various spectra are the durations that were being repeatedly estimated (10 sec, 5 sec, . . . , 0.3 sec). Clearly, the spectra are close to 1=f below about 0.2 Hz. Also, notice that the longer the time interval estimated, the lower the frequencies of the spectra, re¯ecting the duration of the time period over which the experiment continued (see also sec. 15.2). After reading Gilden, Thornton, and Mallon 1995, I reanalyzed some of my own time estimation data in the same way (Ward 1996 and chap. 11). I found similar power spectra near 1=f for time intervals of 1 sec and 3 sec, as well as for ``random'' and self-selected ``preferred'' intervals, even though my data were generated in a di¨erent way. This ®nding is robust. Time estimation is traditionally modeled in psychology by assuming that the intervals bounded by subjects' key presses are produced by the interaction of two processes, an internal clock to time the interval and a motor program to generate the key presses at appropriate moments, both of them also sources of white noise. Gilden, Thornton, and Mallon 1995 changed the traditional model only slightly, by assuming that the internal clock was a source of 1=f noise, whereas the motor program remained a

138

Chapter 16

source of white noise. Figure 16.2B shows the power spectra of simulated time interval generation according to this modi®ed model. The numeric labels on the di¨erent spectra are the ratios of the amount of 1=f noise to the amount of white noise in the respective time series, assuming that the longer the time interval to be estimated, the relatively more 1=f noise contributed by the internal clock. Clearly, the modi®ed model accounts well for the empirical spectra, including the increasing amount of power and the ¯attening of the spectra at high frequencies as the time intervals decrease, both of which re¯ect the increasing prominence of the white motor program noise at shorter intervals. 16.2

Reaction Time

A few months after Gilden, Thornton, and Mallon 1995 appeared, and after informal input by Daniel Gilden about similar experiments he was performing, Keith Clayton presented another signi®cant study (Clayton and Frey 1997) of 1=f noise in human behavior at the August 1995 meeting of the Society for Chaos Theory in Psychology and the Life Sciences. This study was the ®rst reported measurement of power spectra of time series of reaction times, perhaps the most ubiquitous dependent variable used in cognitive science. Despite more than a century of use, however, seldom before had the temporal structure of the variability of reaction times been analyzed. Clayton and Frey (1997) created three conditions of a simple classi®cation task, varying in memory load. In the simplest, subjects pressed one key if a visual stimulus was an ``X''; another key if it was an ``O.'' In the other tasks, subjects had to press one key if the current stimulus was the same as the one on the previous trial, and another if not; similarly, in the most di½cult task, one key if the current stimulus was the same as the stimulus two trials before, and another if not. Clayton and Frey found that the power spectra of the time series of reaction times in the three tasks were di¨erent. The slope of the power spectrum at the lower frequencies for the simplest task was about ÿ0.46; for the one-trial-back, same-di¨erent task, it was ÿ0.35; and for the two-trial-back task, it was ÿ0.26. Thus the slopes varied with memory load, being steepest (most like 1=f noise) for the simplest task and approaching white noise for the most demanding. Interestingly, Clayton and Frey also reported power spectrum analyses of time series formed from iteration of the logistic di¨erence equation (eq. 1.4; see chaps. 1 and 24). They discovered that when the single parameter of the equation equals 3.97, forming a time series that is at the ``edge of chaos'' (see chap.

1=f Noise in Human Cognition

139

24), its power spectrum has a slope of about ÿ0.33 at the lower frequencies. Thus pink noise might even appear in the equations of chaos theory. Gilden (1997) reported similar experiments using several of the most common reaction time tasks, including mental rotation, lexical decision, serial search, and parallel search. Gilden ®tted a version of the same model he and colleagues had used for the time estimation data described earlier to the residual variability (left after subtracting out the mean reaction times) in the time series for each task: Error ˆ …1=f a †n ‡ b
N…0; 1†;

…16:1†

which says that the error on the nth trial is composed of two components: the nth term of a 1=f a noise with mean 0 and variance 1; and a sample from the standard normal distribution, namely, N…0; 1†Ðwhite noise. The constant b determines the relative contributions of the 1=f a and white noises. Gilden found that this model explained over 90% of the residual variance in each of the tasks. The 1=f a noise accounted for from about 0.2 to about 0.4 of the residual variance, the rest being white noise. Moreover, the values of a ranged from 0.7 to 0.9, indicating that the power spectra of the time series of deviations of the reaction times from the mean were near 1=f at frequencies below about 0.1 Hz. Thus near1=f noise characterizes many of the commonest behavioral tasks used in cognitive science laboratories. Impressed by the work of Gilden indicating the ubiquity of the 1=f power spectrum in human behavioral data, and that of Clayton and Frey revealing the possibility that the exponent could be a¨ected by experimental manipulations, I and a colleague (Ward and Richard 2001) decided to investigate speci®c components of the processes giving rise to the 1=f like noise. We had read the description of Schroeder (1991) of a simple set of three ``relaxation''equations that, when summed, would yield a close approximation to 1=f noise. We were aware also of at least three timescales of brain and cognitive processes that might be contributing to such noise; they are summarized in table 16.1. Considering each timescale, t, to be that of a relaxation or memory process, over which activity at a particular moment in time would decay exponentially in its in¯uence on future activity, each of the timescales introduces a correlation scale according to r ˆ e…ÿ1=t† , where r represents the correlation scale; it ranges between nearly 1 (for t approaching in®nity) and nearly 0 (for t approaching 0). A correlation scale value of near 1 means that a given ¯uctuation would in¯uence future ¯uctuations for a relatively long time, whereas a correlation scale near 0 means that a ¯uctuation would in¯u-

140

Chapter 16

Table 16.1 Timescales in the Cognitive and Brain Systems Process (eqs. 16.2 to 16.4)

Timescale t (msec)

Cycle Frequency (Hz)

X

1 to 2.5

Y

Z

Brain System

Cognitive System

Behavioral Task

1,000 to 400

1 to 10 neurons

Preconscious

Peripheral sensory processing

10 to 25

100 to 40

10s to 100s of neurons: within area circuit

Unconscious

Sternberg (1966) STM scanning task

100 to 250

10 to 4

1,000s to millions of neurons: betweenarea loops

Conscious

Crovitz (1970) strobe matching task

ence future ¯uctuations for only a very short time. Schroeder (1991) showed that the sum of three such processes, with timescales spread over at least two decades, is su½cient to approximate 1=f noise. We modi®ed Schroeder's equations to yield the following set: Xn ˆ rx Xnÿ1 ‡ …1 ÿ rx2 † 1=2 Sx Ex; n s

…16:2†

Yn ˆ ry Ynÿ1 ‡ …1 ÿ ry2 † 1=2 Sy Ey; n s

…16:3†

rz2 † 1=2 Sz Ez; n s

…16:4†

Z n ˆ rz Z nÿ1 ‡ …1 ÿ

Wn ˆ …Xn ‡ Yn ‡ Z n † ‡ b
Gn

…16:5†

In equations 16.2±16.4, Xn , Yn , and Z n represent noise generated at three timescales; rx , ry , and rz represent correlations at those scales. They function like the parameters of autoregressive (AR) processes, introducing some proportion of a previous ¯uctuation into each new ¯uctuation, with that contribution decreasing exponentially over the course of the time series (see chap. 11). The S terms, which are our most important addition to the system, are constants that allow us to manipulate at each level separately the amount of uncorrelated white noise, E, relative to that of correlated noise (®rst term). The s term in each equation was added to account for the overall di¨erences in the variance of the time series in di¨erent cognitive tasks; it does not a¨ect the slope of the power spec-

1=f Noise in Human Cognition

141

trum, but only its overall position on the spectral power (vertical) axis. Finally, in equation 16.5, the Xn , Yn , and Z n terms are then combined with more white noise …bGn † to produce each successive time series noise value. Note that equation 16.5 is an expanded version of Gilden's model, equation 16.1, in which the term (Xn ‡ Yn ‡ Z n ) represents the 1=f a process. This modi®ed version, however, provides information about the dynamics previously hidden in the ``1=f black box.'' Also, as I already mentioned, each of equations 16.2±16.4 is basically a ®rst-order autoregressive (ARIMA(1,0,0)) process with parameter rx , ry , or rz , all between 0 and 1, with the term to the right of the plus sign constituting the added white noise (see eq. 11.3). Adding together several ARIMA(1,0,0) processes with a wide range of parameter values turns out to be a very general way to create time series with 1=f -like power spectra and might be a clue as to the necessary and su½cient conditions for such time series to arise (Greenwood and Ward 2001; chap. 19). Although Clayton and Frey (1997) had found that increased memory load decreased the power spectrum exponent a, their task was not the same for all conditions, thus it is uncertain exactly what was a¨ecting the exponent. To ®nd out which scale level was causing the change in exponent, we (Ward and Richard 2001) created a similar decision task in which only the decision load, represented as the number of stimulus and response alternatives, varied between conditions. In our task, subjects either pressed a single key as soon as any one of two short lines appeared in the display (one-choice), pressed one key for a shorter line and a different one for a longer line (two-choice), or pressed the correct one of four di¨erent keys corresponding to one of four lines of di¨erent lengths that could appear (four-choice). In one of several control experiments, subjects had to press a di¨erent, arbitrary, key of the four available on each trial of the task as soon as any line whatsoever appeared, as in the one-choice task except for the rotation among di¨erent response keys. We did this to remove the possibility that any exponent changes might arise solely from the di¨erences between the tasks in the number of motor acts required. Our data were very similar to those of Clayton and Frey (1997), although the power spectrum slopes varied more across conditions, with the onechoice task yielding almost 1=f spectra for several subjects. Figure 16.3A shows the average spectra for each of the conditions across eleven subjects (each subject ran all conditions in partially counterbalanced orders). The slopes for these average spectra were ÿ0.61, ÿ0.36, and ÿ0.24 respectively for the one-, two-, and four-choice conditions, and the linear ®ts (in log-log coordinates) are very good (r 2 ˆ 0:99, 0.97, and 0.96, respectively,

Figure 16.3 Power spectra of reaction time ¯uctuations for 1-choice (1-CRT), 2-choice (2-CRT), and 4-choice (4-CRT) conditions from (A) subject data, (B) simulations of equations 16.2±16.5 with only Sy varying, and (C ) simulations of equations 16.2±16.5 with only b varying. (From Ward and Richard 2001.)

1=f Noise in Human Cognition

143

for the three conditions). The traditional cognitive psychology model that best ®ts our decision task is a Sternberg memory-scanning model, in which some set of remembered representations (1, 2, or 4 of them) must be scanned exhaustively for a match before a response can be made. As table 16.1 indicates, this implies that the middle timescale, at about 10±25 msec per operation, would be most a¨ected by our experimental manipulation because that is the timescale at which the psychological operations involved in Sternberg scanning are said to take place. We did computer simulations of equations 16.2±16.5, varying the contribution of the uncorrelated noise at the middle time scale only (Sy in equation 16.3), to produce the simulated spectra shown in Figure 16.3B. For these simulations we used tx ˆ 1, ty ˆ 10, and tz ˆ 100 so that rx ˆ 0:368, ry ˆ 0:905, and rz ˆ 0:990. The simulations shown in ®gure 16.3B came from runs in which Sy ˆ 1:9 for one-choice, 0.9 for two-choice, and 0.1 for four-choice, with slopes nearly identical to those of the subject data (ÿ0.59, ÿ0.37, and ÿ0.25, respectively) and nearly identical linear ®ts (r 2 ˆ 0:95, 0.97, and 0.95, respectively). The interpretation of this result is that the more decision alternatives there are, the more dominant is the correlation at the middle timescale, causing the spectrum to ¯atten progressively. This interpretation seems reasonable because the correlation at the middle timescale is introduced by the Sternberg memory-scanning process operating at that scale. Varying Sx or Sz also causes changes in the spectra, but the resulting spectra are not nearly linear in log-log coordinates and the slope changes we found are di½cult to simulate. Another possible way to simulate the spectra shown by the subjects is to manipulate only b in equation 16.5, much as Gilden (1997) did. In this case, the contrast is between an una¨ected 1=f process and added white noise, with the white noise process dominating more, the larger b is. Simulated spectra from this manipulation are shown in ®gure 16.3C. In these spectra, b ˆ 1:4 for one-choice reaction times, b ˆ 2:4 for twochoice reaction times, and b ˆ 3:5 for four-choice reaction times, meaning that white noise is more dominant relative to 1=f noise, the greater the decision load. Again, the slopes (ÿ0.60, ÿ0.37, and ÿ0.24) are nearly identical to those in the subjects' data, as are the linear ®ts (r 2 ˆ 0:98, 0.95, and 0.93, respectively). Because in Gilden's model the white noise is supposed to arise from the manual response process, this approach implies that manual response noise plays a greater role the more responses are available, a reasonable statement. Because these spectra also do a fairly good job of mimicking the data, there is some ambiguity remaining as to the best way to model these data. Because, however, our control

144

Chapter 16

one-choice experiment with four arbitrary responses gave roughly the same results as the one-choice experiment with only one response (average power spectrum slope ˆ ÿ0:55), however, we do not think this latter explanation is the best one. Whichever interpretation is eventually accepted, it is clear that there is interesting temporal structure in the ``noise'' that plagues experiments in cognition, and there are indications of how that structure can be revealed. Perhaps future studies, following up on ours, Clayton and Frey's, and Gilden's, will be able to propose and test detailed models in which the modeled cognitive processes naturally give rise to such temporally structured noise.

Chapter 17 1=f Noise in the Brain

[The brain is an] enchanted loom, where millions of ¯ashing shuttles weave a dissolving pattern, always a meaningful pattern though never an abiding one. . . . ÐCharles Scott Sherrington, Man and His Nature The study of systems with strong interactions of many degrees of freedom is one of the most important subjects in physics. ÐE. Novikov et al., ``Scale-Similar Activity in the Brain''

17.1

Neural Activity

Whether a loom or a lake (see chapter 2) is the basis of the metaphorical dynamical brain, our understanding of the dynamics usually begins with its basic functional units, the neurons. As noted in table 16.1, there are many levels of organization of neurons in the brain, from the ion channels in their walls and the second messenger systems in their somata, to the (numerically) huge interconnected networks that insinuate their ®ngers into every area of the entire brain. Although it is surely these many temporal and spatial scales that together make of the brain a special kind of self-organizing system, analysis of brain activity usually begins with the neural action potential level. This is because one of the main ways neurons communicate among themselves and convey information from one place in the brain to another is by sending trains of action potentials along their axons to synapses with dendrites of still other neurons. Each action potential causes the release of transmitter substance across the synapse, stimulating the receiving neuron into ®ring action potentials of its own. Because the action potentials are virtually all identical, the information they convey must be encoded in their time relations. The neuron's instantaneous rate of ®ring (i.e., of generating action potentials) is the dimension usually considered to encode the information, but

146

Chapter 17

more recently the time interval between successive pulses has also been suggested as an encoding dimension (e.g., Singer 1999). Also, synchrony between the ®ring rates of several neurons in a network is supposed to be important, possibly as a solution to the ``binding problem'' of cognitive neuroscience (Engel et al. 1999; see chap. 35). Of course because it will be a¨ected by the ever present noise, the transmission of information by neurons cannot be completely precise. The noise will manifest itself in changes in the rate with which action potentials are generated by the neuron, and possibly in changes in the speed with which each one propagates down the axon. The noise arises from sources such as ¯uctuations in the ionic concentrations inside and outside the axon and other ``channel noise'' (White, Rubenstein and Kay 2000), ¯uctuations in the temperature of the surrounding ¯uids, and the decaying e¨ects of previous action potentials received by the neuron. Some channel noise has been shown to be 1=f , probably arising from the vibration of hydrocarbon chains in the lipids in the nerve cell membrane a¨ecting conductance of potassium ions through the membrane (LundstroÈm and McQueen 1974). The e¨ect of previous action potentials is perhaps the most interesting in the present context, however, because it resembles a memory or relaxation process source for 1=f noise. After each action potential is generated, the neuron experiences an absolute refractory period of about 1 msec, during which no new action potentials can be generated, and an exponentially decreasing relative refractory period of an additional several msec, during which the probability of generating a new action potential gradually increases. Thus the generation of a particular action potential a¨ects the probability of generating another one for quite some time afterwards; the neuron ``remembers'' its previous activity and that memory is combined with current inputs to yield current activity. Musha (1981) did some provocative experiments on the e¨ects of previous action potentials on the time encoding of information by the giant axons of the squid, the easiest of all axons to work with, extensively studied since they were discovered by J. Z. Young. First, by exciting successive action potentials with an electrical pulse, Musha showed that the refractory period also decreases the speed of transmission of the action potential in the axon, dropping from near 25 m/sec for the ®rst excitation to near 10 m/sec for later ones. Clearly, this would a¨ect the encoding of information, whether the instantaneous rate of ®ring or the inter±action potential interval were the encoding dimension. After stimulating the axon with sequences of random electrical pulses (white noise), Musha recorded time series of the ¯uctuations in the time density (the inverse of transmis-

1=f Noise in the Brain

147

Figure 17.1 Power spectra of speed of action potentials in squid giant axons. (From Musha 1981.)

sion speed) of action potentials traveling down the axon. The power spectra for several such time series of density ¯uctuations are shown in ®gure 17.1, along with the power spectrum of the electrical pulses that stimulated the action potentials (at the bottom of the graph). Below about 10 Hz, the action potential power spectra are approximately 1=f , whereas the spectrum of the stimulating pulses in that frequency region is white (¯at). Thus the neurons, the basic building blocks of the brain, themselves display 1=f noise in the foundational mechanism of information transmission, the conduction of action potentials along the axon. Interestingly, Musha and Higuchi (Musha, 1981) had demonstrated the resemblance of the ¯uctuations of action potential speed to ¯uctuations of the speed of automobiles in tra½c. The 1=f ¯uctuations in the tra½c model are attributed to the ``bunching'' of the cars as they are forced to slow down by their proximity to other cars, a property of a more general statistical queueing theory approach to 1=f noise ®rst described by Bell (1960) and summarized in chapter 19. 17.2

Magnetoencephalogram Recordings

From the perspective of physics, the brain is a system with strong interactions of many degrees of freedom. It consists of perhaps 100 billion neurons, each with up to 10,000 connections to other neurons. These connections form hierarchical (and nonhierarchical) groups, from small groups of tens of neurons to large groups consisting of entire sensory, cognitive, or motor processing areas, such as the visual cortex, with many

148

Chapter 17

Figure 17.2 Positions of magnetoencephalogram (MEG) channels. (From Novikov et al. 1997. ( 1997 by The American Physical Society.)

millions of neurons. In physics, such systems are usually described by a ``similarity regime'' (see sec. 18.3), in which similar behavior is observed at several scales. Under certain conditions, a 1=f power spectrum of temporal ¯uctuations can arise from such a similarity regime. Novikov et al. (1997) recorded the magnetoencephalogram to establish the existence of such a regime in the human brain. To do this, they had to use a superconducting quantum interference device (SQUID), which allowed them to record the magnetic ®eld at the surface of the brain generated by the ¯ow of electrical current in and around neurons. They placed sensors in an array around the heads of their two human subjects and had them lie down in a relaxed way. The placement of the sensors is shown in ®gure 17.2. Then they recorded the magnetic ®eld strength at each sensor location at 231.5 Hz (one record every 1=231:5 ˆ 0:0043 sec) for 30 min to form time series for analysis. Figure 17.3 shows a representative power spectrum from one of the sensors for each of the two subjects (left, female; right, male). These

1=f Noise in the Brain

Figure 17.3 Power spectra of magnetoencephalogram (MEG) time series. (From Novikov et al. 1997. ( 1997 by The American Physical Society.)

149

150

Chapter 17

spectra show peaks at a few speci®c frequencies (about 1.5 and 4 Hz for the female; about 10 Hz for the male) re¯ecting endogenous brain rhythms such as the alpha wave (see sec. 17.3). More interesting are the average slopes of these power spectra, represented by the straight lines in the graphs. These lines have slopes of ÿ1.03 and ÿ1.19, respectively over the range 0.4 to 40 Hz, very near the ÿ1 expected for 1=f ¯uctuations. Power spectra for other sensors were very similar. Moreover, taking the di¨erence between pairs of sensors usually eliminated the peaks and yielded even more stable 1=f spectra. The average spectrum slopes for the two subjects over all pairs of sensors were ÿ0.98 and ÿ1.28, for the female and the male, respectively. Further analyses of the data established that the scale similarity implied by the 1=f spectra was relatively ``local,'' meaning that it extended over only limited brain areas, probably related to the shared function of those areas. Thus the spontaneous activity of functionally related chunks of the human brain exhibits 1=f noise. This ®nding supports the assumptions made in chapter 16 regarding the origin of the 1=f spectra in human cognition. 17.3

Electroencephalogram and Event-Related Potential Recordings

A less exotic and less expensive way of noninvasively studying brain activity is to record the electrical ®eld at the scalp generated by the electrical current ¯owing around and through the neurons when they generate action potentials (see Picton, Lins, and Scherg 1995). This recording is called the ``electroencephalogram'' (EEG), and its study has yielded much information about brain activity in humans. Among other things, common EEG brainwaves, such as the alpha wave can be used to classify mental illnesses (e.g., John 1977) and the stages of sleep (e.g., Dement 1974). When the EEG is used to record brain potential ¯uctuations over shorter intervals in response to particular sensory stimuli, the recording is called an ``event-related potential'' (ERP). Cognitive neuroscientists use ERPs extensively to characterize the brain activity related to stages of information processing in humans (see chap. 33). However, as might be expected, ERPs are small signals (often < 2 microvolts, or mv) buried in noise (often > 60mv). Thus the typical ERP study averages ERPs from many trials, taken in a sequence, and thus achieves a relatively clear and noise-free picture of the brain activity that occurred on each of the 50±100 or more trials on which a given stimulus was presented. Although it seems di½cult if not impossible to imagine obtaining a time series of single ERPs that could be investigated with the fast Fourier transform (FFT), I and a col-

1=f Noise in the Brain

151

Figure 17.4 Typical event-related potential (ERP) at Cz location to a sound stimulus. (From McDonald and Ward 1998.)

league (McDonald and Ward 1998) did just that; schemes are being proposed (e.g., JasÂkowski and Verleger 1999) to remove at least some of the background noise from recorded individual ERPs in a more general way. In our study, MacDonald and I used an ERP that was larger than most and very robust, one evoked by the presentation of a 50 msec, 1,000 Hz pure tone at 80 dB to a human subject seated in a very quiet (35 dB background noise) sound-attenuating room. In this context, the tone sounds fairly loud, and the ERP it evokes is quite large. Figure 17.4 shows one typical average ERP to this tone, averaged over 1,000 presentations. Several things should be noted about such ERPs before we consider their temporal distribution. First, the graph displays the electrical voltage recorded at electrode site Cz, in the center of the top of the head, the best site at which to record the activity in the auditory cortices. At Cz, the electrode records electrical ®elds from both auditory cortices, located in the left and right temporal lobes of the brain, just above the ears, as well as from other brain areas. The electrical voltage at Cz was measured in reference to another electrode site, in this case a site just over the mastoid bone at the base of the skull behind the ear. In ®gure 17.4, the electrical voltage is displayed as it varied over time from about 100 msec before the tone was presented to about 800 msec after the presentation began. Second, the prominent neural activity components labeled P1, N1,

152

Chapter 17

Figure 17.5 Power spectra of event-related potential (ERP) time series: (A) 5 msec before end of prestimulus baseline; (B) N1 peak amplitude. (From McDonald and Ward 1998.)

1=f Noise in the Brain

153

and P2 in ®gure 17.4 correspond to neural activity in the auditory cortices and elsewhere in the brain at di¨erent intervals after the tone was presented, Third, these components are quite large, for example, the N1 component deviates about 5 mv from baseline (the period to the left of the vertical axis at 0 msec), making them good candidates for a time series FFT analysis. Finally, the EEG during the baseline period is quite ¯at, indicating that the auditory cortex showed no systematic activity just before the onset of the tone. McDonald and I recorded time series of 1000 ERPs from tones presented every 2 sec under two conditions, in one of which the subject ignored the tones and in the other of which he counted them. This made only a small di¨erence to both the average ERPs (®gure 17.4) and the FFT analyses, and so I will present only the count condition results. We decided to look at the time series formed from two points on the ERPs, one 5 msec before stimulus onset and the other 96 msec after it, at the average location of the N1 peak. We did FFTs on the time series formed from the 1,000 successive ERPs taken 2 sec apart at each of these points. Figure 17.5 shows the resulting power spectra, one from the N1 peak and one from the prestimulus period. Both are approximately 1=f and are typical of all of our results. Not only is the evoked activity of the brain 1=f , but also the noise in which the evoked activity is embeddedÐand which possibly characterizes the process generating that evoked activityÐ is also 1=f . It is clear that the human brain is characterized by 1=f noise in many activity regimes, from ion ¯ow in neurons to activity evoked by external stimuli. The challenge now is to discover the functional implications of this fact.

This page intentionally left blank

Chapter 18 Models of 1=f Noise

The author has for some time believed that the inverse-frequency spectrum must be due to some form of co-operative phenomenon . . . that it is a property of the behaviour of the whole complex of electrons in the conduction band, and not a property of the mechanisms associated with the individual electrons. ÐD. A. Bell, Electrical Noise

18.1

The Simplest Case

The simplest, most direct way to produce a power spectrum de®ned by the equation S… f † ˆ 1=f a

…18:1†

is to add together a set of periodic (sine and cosine) functions with the frequency amplitude weighted according to equation 1.1 and random relative phase. Alternatively, we can produce the required power spectrum indirectly, by ®ltering a broadband noise through a low-pass ®lter with a shoulder that falls o¨ as 1=f a (see Corsini and Saletti 1987). However, because they merely reproduce the de®nition, equation 18.1, these approaches seem like cheating. What is needed is a description of how colored noise, and in particular 1=f noise (a ˆ 1 in eq. 18.1), arises naturally and inevitably from the cooperative activity of the elements of dynamic systems (possibly at the ``edge of chaos''; see Kau¨man 1993). Outlined below are a few of the most compelling accounts of how 1=f noise can arise in dynamic systems. (One not covered in this chapter, but discussed in chapter 16, is the presence of correlations in a time series at several timescales covering at least two log units; others not covered here or in chapter 19 can be found mostly in the physics literature; just search for ``1=f .'')

156

Chapter 18

18.2

Multiplicative Noise

Remember from chapter 14 that, when noise is added to a dynamical equation (the Langevin approach) as in equation 14.2, the noise simply blurs the dynamics. Caused by some force acting on the dynamical system from the outside, such noise is called ``external noise'' (van Kampen 1987). Although in principle it can be switched o¨, in practice this is seldom possible. In contrast, multiplicative noise is internal noise, arising from the dynamical system itself. Being a part of the system's dynamics, internal noise cannot even in principle be switched o¨; it is a collective or systemic phenomenon arising because matter consists of particles or because energy is quantized. The simplest way to model internal noise is with a system in which such noise constitutes the only dynamics. An example of this approach was created by Thurner, Feurstein, and Teich (1997). They used a system composed of a set of variables wi …t†, i ˆ 1 to N, where the value of each variable at time t ‡ 1 equals its value at time t multiplied by a noise term l i : wi …t ‡ 1† ˆ l i wi …t†;

…18:2†

and where l i is a sample from a probability distribution. The values of l i are independent and identically distributed (i.i.d.; see chap. 14). The system of N variables is ``coupled,'' in other words, a correlation between their values is introduced, by requiring that N X

wi …t† ˆ N:

…18:3†

iˆ1

Thus, on each iteration of the model, after equation 18.2 has been applied to fresh samples of noise, the new values of wi …t† are multiplied by X wi …t†; N= i

implementing equation 18.3. Notice that this also makes the average value of the variables equal to 1 because, after the transformation, X wi …t†=N ˆ N=N ˆ 1: i

The model also requires that wi …t† > w0 b 0;

…18:4†

so that the noise tends to make all of the variable values larger, a tendency that is held in check by the application of equation 18.3 on each

Models of 1=f Noise

157

iteration. Notice that, in this model, all changes in variable values are induced by the multiplicative noise. That is, without the in¯uence of the multiplicative noise, there would be no dynamics at all. Thurner, Feurstein, and Teich (1997) simulated the temporal unfolding of this multiplicative noise system on a computer, drawing the l i from a uniform distribution and recording the successive values taken by each of the N variables wi …t† at each time point t. They then calculated the fast Fourier transform (FFT) of the time series for a representative individual variable (i.e., an arbitrary value of i) and used it to obtain a typical power spectrum for the system. This spectrum approached 1=f as the lower bound w0 approached zero, and was approximately ¯at (white noise) for w0 > 0:8 and N > 50. In addition, the occurrence frequencies of values taken by the variables wi …t† over a large number of time steps were a power function of the values themselves, with the exponent depending on w0 , the other side of the self-organized criticality explanation of the 1=f power spectrum (see sec. 18.3). A similar but more complicated approach is to combine additive and multiplicative noise according to the following general di¨erential equation: dx ˆ l…t†
x…t† ‡ h…t†; dt

…18:5†

where x…t† is the system variable, l…t† is the multiplicative noise, and h…t† is an additive noise (see Nakao 1998). The meaning of the di¨erent terms of equation 18.5 varies with the situation. For example, in an economic model, x…t† could represent the wealth of a country; l…t†, the (noisy) rate of change of wealth from internal production dynamics; and h…t†, noise from outside sources such as the world economy, wars in other countries, and so forth. Nakao (1998) and others have demonstrated that this model, which can be applied to many situations, both physical and nonphysical, does reproduce the prototypical power law of amplitudes and the associated 1/f power spectrum of systems characterized by selforganized criticality (see sec. 18.3). 18.3

Self-Organized Criticality

What do sandpiles, the game of life, earthquakes, turbulence in liquids, and stock prices have in common? According to Per Bak (1990), they are all examples of self-organized criticality, whereby a dissipative dynamical system with many degrees of freedom operates far away from equilibrium

158

Chapter 18

Figure 18.1 Self-organized sand pile. (From Bak 1990. ( 1990. Reprinted with permission from Elsevier Science.)

near a point of instability, called ``criticality''; because it does so by itself, without any ®ne-tuning by an external driving in¯uence, the system is said to be ``self-organized.'' At the point of instability, small ¯uctuations can give rise to events of any size, and the likelihood, or probability density, of any particular size of event, D…s†, is a power function of the event size, s: D…s† ˆ ksÿt :

…18:6†

In other words, the larger the event (the more elements it involves), the less likely it is to occur, although if we wait long enough, eventually an event of any size (up to the system size) will occur. For Bak (1990), the prototype of a self-organized critical system is a model of a sandpile created by dropping sand grains one at a time onto a particular location (®g. 18.1). As the pile grows, avalanches of various sizes occur until eventually the sides of the pile settle into a particular ``critical'' angle, at which the pile ``is barely stable with respect to small perturbations'' (Bak, Tang, and Wiesenfeld 1987, 382). As the sand grains continue to be dropped on the pile, avalanches of all sizes (up to the size of the pile) occur with a probability distribution like that described by equation 18.6. The sandpile looks the same no matter at what scale it is viewed, from the scale of the individual grain of sand to the scale of the movement of many grains in large avalanches. Such systems are called ``self-similar.''

Models of 1=f Noise

159

How, then, do self-organized critical dynamical systems give rise to 1=f noise? According to Bak, Tang, and Wiesenfeld (1987), they do so inevitably in the following sense. First, the lifetime of an event in the system, t, is related to the size of the event, s, by the equation t 1‡g A s, or, solving for t, t A exp‰log s=…1 ‡ g†Š, where g is the rate at which the event propagates across the system elements. For example, in the sandpile, t would be the lifetime of an avalanche; the larger the avalanche, the longer the lifetime because the larger avalanches take longer to occur, involving as they do energy transfers between many elements over large distances. Given this relation, the distribution of lifetimes can be calculated from the distribution of event sizes:   ds …18:7† D…t† ˆ …s=t†
D…s…t††
A tÿ…g‡1†t‡2g 1 tÿb dt The power spectrum of the probability density function of lifetimes (eq. 18.7) is given by … …18:8† S… f † ˆ ‰…t
D…t††=…1 ‡ … f
t† 2 †Š dt A f ÿ2‡b ; which states that the power spectrum is a power function of frequency, f , with exponent equal to ÿ2 ‡ b (you can ignore the integral expression in the middle if you ®nd it puzzling). Thus a power law for the size distribution of events in a system exhibiting self-organized criticality (eq. 18.6) implies a power spectrum for the temporal ¯uctuations that is 1=f like. In fact, we can calculate the exponent of the power spectrum from the exponent of the size distribution, t, and the rate of event propogation, g, by plugging the expression for b from equation 18.7 into the expression for the exponent in equation 18.8. Using a for the exponent of the power spectrum function of f , as usual (eq. 18.1), we have a ˆ ÿ2 ‡ ‰…g ‡ 1†t ‡ 2gŠ. In a computer simulation example given by Bak, Tang, and Wiesenfeld (1987), t was measured to be 0.98, and g was about ÿ0.188, which means that b was about 0.42 and the slope of the power spectrum …ÿ2 ‡ b ˆ ÿ2 ‡ 0:42† was ÿ1.58. In other words, S… f † A 1=f 1:58 . (Actually, Bak, Tang, and Wiesenfeld measured b directly and thus did not have to worry about g; I inferred it from their numbers.) In another case, they obtained b ˆ 0:90 and S… f † A 1=f 1:1 . More recently, Maslov, Tang, and Zhang (1999) reported 1=f power spectra for the dynamics of model sandpiles con®ned to narrow stripes (almost one-dimensional). For these models, however, the 1=f power spectrum was not related to power

160

Chapter 18

law distributions of avalanche sizes: the constraints on the sandpiles resulted in the almost continual avalanches making only small changes to the con®guration of sand grains. The memory of the system for previous con®gurations was nonetheless quite long and led to the observed 1=f power spectrum. 18.4

Center-Surround Neural Network

This model, based on the center-surround organization of neural networks in sensory systems and elsewhere in the brain, is perhaps the one most directly relevant to dynamical cognitive science. Usher, Stemmler, and Olami (1995) simulated a type of neural network that exhibits 1=f ¯uctuations in the temporal dynamics of individual units, although its overall behavior is a type of self-organized criticality. They created a 100  100 lattice of model neurons, with each functioning like an integrate-and-®re neuron (see chap. 22), that is, each neuron summed exponentially decaying (20 msec time constant) voltage inputs and ``®red'' a pulse if the total voltage reached a threshold. After ®ring, the neuron voltage was reset by subtracting the threshold. Each model neuron received excitatory inputs (pulses) from N other neurons in a disklike region around it, and inhibitory inputs from a ring of N neurons outside the disk. In other words, each neuron modeled the center-surround organization found in many regions of the brain, especially in sensory cortex (see, for example, Zeki 1993). In addition, each cell received external noise in the form of Poisson-distributed excitatory pulses of magnitude 1/N arriving at a rate lP . The relative weights of the excitatory and inhibitory inputs were WE ˆ a=N and WI ˆ bWE . Of course the behavior of this network varies with the values of the parameters a; b, and lP . When b, the relative weighting of inhibitory inputs to each neuron, is a substantial fraction of a, the relative weighting of excitatory inputs, for example a ˆ 1:3 and b ˆ 0:67, clusters of high ®ring activity develop and move around the lattice. Figure 18.2 shows the path of one such cluster over successive 50 msec intervals (1 msec is the simulated cycle time of the network, activity is summed over 50 msec intervals to yield the clusters). Interestingly, the behavior of these clusters varies as a function of the external noise input lP . When lP is relatively small, the clusters move randomly, but as lP is increased, the movements take on a pattern, becoming unidirectional at high values of lP , which results in diagonal stripes in the long-term average activity picture. At

Models of 1=f Noise

161

Figure 18.2 High-activity clusters move on the center-surround lattice. (From Usher, Stemmler, and Olami 1995. ( 1995 by The American Physical Society.)

intermediate values of lP , the movement ceases and the clusters show a rigid spatial structure. Where, then, is the 1=f noise in all of this? It turns out that the behavior of single units in the lattice exhibits the required ¯uctuations, but only for certain parameter values. First, the values of a and b must be appropriate for producing the slowly moving clusters shown in ®gure 18.2. Then the value of lP must be such that the clusters are in the transition region between ``¯uid'' (random movement) and ``crystalline'' (no-movement) regimes. Under these conditions time series of the individual units' ®ring rates show near-1=f ¯uctuations. Now, apparently, under these same conditions changes in the value of b also change the responsiveness of the model neurons to ``stimuli,'' in this case small changes in the external input rate lP . Figure 18.3 shows the mean spike rate per millisecond for a region of the lattice of model neurons where the external

162

Chapter 18

Figure 18.3 Responses of model neurons in various parameter regimes to stimuli; note priming e¨ect in inset. (From Usher, Stemmler, and Olami 1995. ( 1995 by The American Physical Society.)

input rate to the excitatory center neurons of the region was raised by 12% for a brief interval. Such a stimulus, is equivalent to a brief change in the mean amount of energy impinging on a sensory receptor, for example, a ¯ash of light to the retina of the eye. For lower values of b, which result in relatively less inhibition from the surrounding inhibitory neurons, the system is more ¯uid (and the power spectrum more like white noise), and the model neurons exposed to the stimulus give a very strong but quickly decaying response, which indicates no ``memory'' (top trace of ®g. 18.3). For higher values of b, which result in relatively more inhibition from the surrounding neurons, the system is more rigid or crystalline (and the power spectrum more like black noise), the stimulated neurons respond only weakly to the stimulus, and the weak response decays only slowly to baseline, indicating perhaps too much ``memory'' even for a weak response (bottom trace). For the value of b that yields 1=f ¯uctuations, however, the response from the stimulated neurons is both vigorous and has just the right amount of memory, like Baby Bear's porridge, decaying to a level close to baseline over an interval of about 100 msec (middle trace). Priming is said to have occurred when presentation of a stimulus facilitates the response to a subsequent stimulus presentation (e.g., Posner 1978), as demonstrated by the response to two successive stimuli illustrated in the inset of ®gure 18.3, where the response to the second stim-

Models of 1=f Noise

163

ulus, which occurs while the response to the ®rst is still decaying and is relatively far from baseline, is even more vigorous than the ®rst response. Priming is a ubiquitous psychological phenomenon observed in many paradigms, from memory to attention to perception and even possibly subliminal perception (e.g., Marcel 1983). That this neural network exhibits priming only in the 1=f parameter regime implies that there might be something special about that regime in neural computation and in the functioning of real brains.

This page intentionally left blank

Chapter 19 Statistical Theory of 1=f Noise

The simplicity of the 1=f spectrum encourages speculation . . . that some general statistical considerations may account for it. . . . It appears that the prevalence of 1=f noise is nearly equivalent to the assertion that distribution functions for logarithms of rates tend not only to be broad but to have wings that fall o¨ much less sharply than Gaussians. ÐM. B. Weissman, ``Survey of Recent 1=f Theories'' Di¨erentiating a logarithmic cumulative distribution results in a 1=f spectral density. . . . ÐStefan Machlup, ``Earthquakes, Thunderstorms, and Other 1=f Noises'' One cannot understand a 1=f noise without searching beyond its spectrum and the search for a single explanation becomes hopeless. ÐBenoit Mandelbrot, Multifractals and 1=f Noises

19.1

What Must Be Explained

The characterization of 1=f noise depends on the concept of a power spectrum, which is the square of a Fourier transform of a time series, a correlation function, or a probability distribution arising from some kind of a random process. The temporal structure of the stochastic process is revealed through the distribution of frequencies of deterministic periodic functions of time. Although this idea is accepted today, it was not always. In his excellent discussion of electrical noise, Bell (1960) mentioned that at one time it was seriously questioned whether a time-varying random process could properly be said to possess a power spectrum, or even a Fourier transform. Because a mathematical technique can be applied wherever the conditions for its application exist, a Fourier transform of any function can be obtained. The result may not have any meaning, however, and this is what was questioned. It is now clear that not only does the Fourier

166

Chapter 19

transform of a temporal random process exist, it also has a meaning that is similar to, although somewhat more abstract than, that of a Fourier transform of any other time-varying function, with obvious implications for the existence of the power spectra of random processes. In fact, the concept of the Fourier transform applies strictly only to mathematical functions and all such are abstractions. Any application to real measurements is only approximate, almost always through the fast Fourier transform (FFT), although if the measurements are exquisitely ®ne and frequent, the approximation can be very close. Thus a random process does have a meaningful power spectrum, even if it is only ¯at, as in white noise. Another problem with the characterization of 1=f noise is that the 1=f power spectrum can be strictly true only over a very limited range of frequencies; there must be both high- and low-frequency bend points. The high-frequency bend point exists because Planck's constant determines a scale, the quantum scale, for both length and time (the Planck length and the Planck time). Above this point, the power spectrum of a 1=f noise becomes 1=f 2 (e.g., Machlup 1981). The low-frequency bend point exists because of the Nyquist limit: the measurement time must be at least the inverse of the frequency (e.g., to measure noise at 0.1 Hz, measurement time must be at least 10 sec). But in practical situations, the limits are even narrower than this. Measurement times are usually severely limited, and only the most esoteric (and expensive) experiments reach even halfway to Planck limits. Thus we are usually looking at power spectra that are 1=f -like over only a few decades (powers of 10) of frequency, sometimes only over one decade. Moreover, the spectra involved are usually obtained via the FFT, as mentioned earlier, the square of which measures the power spectral density in a small set of rather large (e.g., 1/32 octave), discrete frequency bands. Thus, although pure theory often stumbles over technical issues, the need for explanation is actually for a much less demanding set of facts. What must be accounted for is (1) that the power spectrum of a time series is 1=f -like over one or more decades of frequency, and (2) that such power spectra are so ubiquitousÐthey appear far more often in time series than would be expected on the basis of ``ordinary physics'' or even of ``ordinary cognitive science.'' 19.2

Queuing in a Wire

Scientists have been searching for a more general explanation of 1=f -like noise for many years, and it was early recognized that some general sta-

Statistical Theory of 1=f Noise

167

tistical considerations might be deeply implicated (see Weissman epigraph above). Bell (1960) was one of those who searched, and he proposed an ingenious theory, based on statistical queuing theory, of how 1=f noise arises in current ¯uctuations in an electrical circuit. In Bell's theory, as is common in physics and electronics, electrical current is conceptualized in terms of the ¯ow of carriers (either electrons, which have negative charge, or ``holes,'' which have positive charge) through a wire. Let us think in terms of electrons in a copper wire. The electrons can exist in an electron shell around an atom in a bound stateÐor in an excited, unbound state, where they are free to move through the wire with other excited electrons in what is called the ``conduction band.'' The electron is excited from the bound state by the absorption of a quantum of energy (a photon). After it spends some time in the conduction band, participating in the current ¯ow, it can fall back into the bound state by emitting a photon and joining the other electrons bound to a particular electron shell around a speci®c atom or molecule. This process repeats itself at random times throughout the duration of the current ¯ow. In other words, the electrons oscillate randomly between bound and excited states. It is some aspect of this process that generates ``current noise,'' which has a 1=f spectrum over many decades of frequency. Bell (1960) conceptualized the problem of explaining current noise in a di¨erent way from the usual. Instead of asking how long, on average, a carrier stays in the conduction band, he asked, ``If a carrier has once been excited into the conduction band, how long is it likely to remain there before it (rather than its neighbors which may have arrived in the conduction band earlier or later) chances to fall back to a bound level.'' (Bell 1960, 231) From this point of view, the problem appears to be one of congestion: for a given electron to fall back into a bound state, there must be a ``vacancy'' in the vicinity. The vacancy the electron left when it moved into the conduction band may not be available, either because it moved away or because some other electron now occupies its former home. Moreover, wherever it is, there is likely to be a lineup (or queue) for the available vacancies in the bound state (available electron shells with fewer than a full complement of electrons) because, by chance, electrons may have bunched up at a particular place in the wire (even though other vacancies may have no lineup and may stay vacant for a long time). Thus, even though the average time an electron spends in the conduction band may be only a few microseconds, the ¯uctuations of the number of electrons in the conduction band, which are the source of current noise, can exhibit periodicities over tens of thousands of seconds. Figure 19.1

168

Chapter 19

Figure 19.1 Cartoon of carriers (squares) queuing in a wire.

shows a cartoon of this situation: the circles represent copper atoms with (for illustrative purposes only) two electron sites (short lines on the circle) per atom. If an electron site is occupied by an electron (small square), then that site is not available for any other electrons to occupy. The atom in the center of the diagram is shown with several electrons clustering around it. None of these electrons can occupy either site in the atom until one of the electrons in one of the sites leaves the site vacant by joining the conduction band. Thus there is a queue of electrons in the conduction band waiting for a vacant site in that area of the wire, even though there are vacant sites in another atom some distance away (at the left of the wire). The statistical description of this queuing situation, called ``queuing theory,'' has been known for some time (see Feller 1968, 1971). Bell (1960) was able to (loosely) map the situation of electrons ¯owing in an electrical current through a wire onto a particular variation of queuing theory. In particular, he was able to discover an expression for the probability distribution of the time an electron spent waiting in a queue. Each probability density in the density function was a summation of a set of exponential distributions that had a time constant ti . Bell argued that the various exponential distributions in the sum should be equally weighted because each had an equal probability of occurring. It was known that a time series generated by summing samples from such an equally weighted set of exponential distributions with widely varying time constants leads to a power spectrum that is proportional to 1=f , whereas the power spectrum of a time series formed from samples from a single exponential distribution is proportional to 1=f 2 . Bell therefore concluded that the 1=f power spectrum of current noise arises from the statistical properties of the electrons' queuing behavior while in the conduction band. He also

Statistical Theory of 1=f Noise

169

showed that various departures of power spectra from 1=f , in particular, the dependence of the slope of the power spectrum on the particular conductor or semiconductor (the material from which transistors are made), can be explained by variations in how the summation arises or in the limits over which it occurs. Bell's approach (1960) was one of the ®rst to o¨er a general explanation for why so many real dynamical systems exhibit 1=f -like power spectra. It is su½cient that some process of the system generate an equally weighted set of exponential distributions with a wide range of time constants that all sum to produce at least a part of the time-varying behavior of the system. As an interesting corollary, this process is most likely to be a cooperative phenomenon, one that results from the interactions of all of the system's elements (e.g., the queuing of carriers in a wire), rather than from the actions of individual elements. Indeed, it may be useful to think of it as a queuing process in a very general sense, although there are other stochastic processes that could result in the required statistical description of temporal behavior. For example, a collection of neurons could generate 1=f -like ¯uctuations because of the action of reverberatory circuits at several timescales, as described in section 16.2: the necessary exponential distributions with di¨erent time constants would be generated as circuits of di¨erent sizes ``relaxed'' or ``forgot'' their response to a particular input. The larger the number of neurons involved (or entrained) in the circuit, the more slowly it would relax to the unperturbed state (the larger would be the time constant for that distribution). In such a case, queuing theory would not describe the route to the required set of exponential distributions, and the relaxation approach would lead to a better mechanistic understanding. 19.3

ARIMA(1, 0, 0)

Another interesting aspect of the example explored in chapter 16 is that the set of exponential distributions need not be complete for an approximation of a 1=f spectrum to exist. In that example, only three relaxation processes are involved and the approximation is quite close. Moreover, because the three are ARIMA(1, 0, 0), or AR(1), processes (as hereafter abbreviated; see chap. 11), that is, linear autoregressive processes, the dynamical processes need not be nonlinear. The most important requirements of the set of distributions involved are (1) that there be several of them; (2) that their time constants be spread over at least two decades;

170

Chapter 19

and (3) that they be more or less equally weighted. These three requirements correspond to those stated in chapter 16: several widely spread timescales and no one preferred timescale. Thus what I am asserting here is that any process giving rise to these properties will exhibit 1=f -like ¯uctuations. Dynamical systems made up of many complexly interacting elements will often exhibit this kind of ¯uctuation because of the cooperative or emergent behavior of the elements in producing the dynamical behavior of the system (see chap. 6). The challenge is to show, in any particular instance, just how the necessary exponential distributions and time constants arise. Several authors have argued that particular combinations of linear processes can be responsible for 1=f -like behavior. Milotti (1995) developed a modern version of Bell's model (1960) of current noise that depends only on linear processes and some very general mathematical properties of them. Hausdor¨ and Peng (1996) showed that the simple linear superposition (summation) of random processes acting on several timescales, in a process they called ``multiscaled randomness,'' could lead to 1=f -like behavior similar to that found in many biological systems. Granger (1980; Granger and Ding 1996), an economist, was perhaps the ®rst to suggest aggregating AR(1) processes to produce time series exhibiting 1=f -like behavior. His work has been developed further by mathematicians (e.g., Leipus and Viano 2000). Finally, Kaulakys and MesÆkauskas (1998) and Kaulakys (1999), apparently unaware of the earlier work, proposed using events whose recurrence times obey an AR(1) process to produce 1=f noise. In Greenwood and Ward 2001, we attempted to generalize this AR(1) approach. As noted in equation 11.2, an AR(1) process is written Yt ˆ f1 Ytÿ1 ‡ at ; where here 0 < f1 < 1 and at is a sample from a white noise process. If f1 ˆ 0, then the process just produces a time series of samples of white noise, which has a ¯at power spectrum, whereas if f1 ˆ 1 the process is an integrative process, for example, a random walk, and produces a time series with a power spectrum proportional to 1=f 2 . Consider that the ¯at power spectrum of a time series of white noise samples is equivalent to a spectrum proportional to 1=f 0 . Thus, as others have also noticed, the 1=f power spectrum is in some sense between the power spectra produced by these two limiting AR(1) processes. We speculated that perhaps the power spectra of AR(1) processes with parameter values between those producing the limiting power spectra would be close to 1=f , or at least 1=f -like.

Statistical Theory of 1=f Noise

171

Although the power spectra of time series generated from individual AR(1) processes with 0 < f1 < 1 are not completely linear in a log-log plot, they do have linear sections at the higher frequencies, and the slopes of these sections are between 0 and ÿ2, approaching these numbers as limits as the parameter value of the AR(1) process approaches 0 or 1, respectively. Importantly, when several such AR(1) processes, with parameters su½ciently far apart, are added together to form a single time series, the entire power spectrum becomes roughly linear in a log-log plot, re¯ecting the 1=f -like behavior of the time series. Figures 19.2A and 19.2B reiterate the point that only three such processes are su½cient (notice that, again, we have averaged groups of points at higher frequencies in these spectra, as recommended by Mandelbrot 1998). Indeed, the more such processes are added together, the ¯atter the power spectrum becomes. Figure 19.2A was created from FFTs of 100 separate realizations of time series 1,024 observations long, each made up of the sum of three AR(1) processesÐwith white noise sampled from N…0; 1† and with parameters (f1 ) of 0.1, 0.5, and 0.9. Figure 19.2B shows an analytical calculation of the same power spectrum, obtained by adding together the calculated power spectra of the three separate AR(1) processes: Stotal … f † ˆ S0:1 … f † ‡ S0:5 … f † ‡ S0:9 … f †;

…19:1†

where the power spectrum for each separate process with parameter fi was calculated from Si … f † ˆ

2p…1 ‡

fi2

1 ; ÿ 2fi cos f †

…19:2†

taken from Brockwell and Davis 1996. Clearly, the summated AR(1) process gives a good approximation to a 1=f power spectrum (slopes of both straight lines are ÿ1.03), in this case, across about two decades of frequency. Thus the simple aggregation of linear random processes at several scales can give rise to a time series with a 1=f power spectrum (cf. Hausdor¨ and Peng 1996). 19.4

Multifractals and Wild Self-A½nity

In his several works on multifractals, Benoit Mandelbrot (e.g., 1998) has made arguments similar to those above. Pointing out the several mechanisms by which 1=f -like power spectra can arise, he evocatively described the exotic and subtle manifestations of such mechanisms, in particular, ways in which the time series arising from di¨erent mechanisms di¨er,

172

Chapter 19

Figure 19.2 (A) Approximate 1=f power spectrum (mean G 1 sd) from a Monte Carlo simulation of the sum of three ®rst-order autoregressive [AR(1)] processes; (B) approximate 1=f power spectrum calculated for the sum of three AR(1) processes. (From Greenwood and Ward 2001.)

Statistical Theory of 1=f Noise

173

despite their similar power spectra. However, I disagree with his conclusion that there is no real commonality between these mechanisms except that revealed by their power spectra. The statistical considerations described above provide a unifying theme that goes beyond the mere speci®cation of a 1=f power spectrum. In particular, the statement that cooperative systemic behavior leads to 1=f -like ¯uctuations seems to me to be a powerful explanatory device. A marriage of general systems theory with ``complexity theory'' (which relies on discovering 1=f power spectra but seldom goes further) promises new directions for both. Mandelbrot's main point (1998) is that the 1=f power spectrum is only one of a set of symptoms for what he holds to be a more fundamental concept, self-a½nity, which requires that the graph of the function specifying a process be invariant under a set of linear transformations (i.e., stretching or shrinking along each axis). Mandelbrot argues that, for the power spectrum of a noise to be proportional to f ÿa , it is su½cient that the noise be self-a½ne (and also not too weird). Because he has experienced no case of 1=f noise that is not self-a½ne, he uses ``1=f noise'' and ``self-a½nity'' interchangeably, although he believes that the two terms do not have the same explanatory status: ``The analytic term 1=f noise is misleadingly speci®c, while the geometric term self-a½nity is realistically general'' (81). If self-a½nity and f ÿa power spectra imply each other (see Greenwood and Ward 2001), then either can be seen as fundamental. However, as is his wont, Mandelbrot opts for the geometric description as more revealing and more fundamental. This point of view resembles that of the promoters of chaos theory (see chaps. 23±27). It reveals a tension between the classical way of doing science, which depends heavily on analytic descriptions of phenomena and theories, and the many competing new paradigms, such as chaos theory, complexity theory, and Mandelbrot's self-a½nity theory, which often depend on geometric arguments and descriptions. Such descriptions can be less satisfying than classical ones, although sometimes they are all that is possible (see Kellert 1993). At any rate, Mandelbrot (1998) provides many tools for the further speci®cation of wildly self-a½ne processes in terms of the mechanisms from which they arise. Such speci®cation is surely of value any time the mechanism of a 1=f -like noise is to be explained.

This page intentionally left blank

Chapter 20 Stochastic Resonance

Noise in dynamical systems is usually considered a nuisance. But in certain nonlinear systems, including electronic circuits and biological sensory apparatus, the presence of noise can in fact enhance the detection of weak signals. This phenomenon, called stochastic resonance, may ®nd useful application(s). . . . ÐKurt Wiesenfeld and Frank Moss, ``Stochastic Resonance and the Bene®ts of Noise''

20.1

What is Stochastic Resonance?

Stochastic resonance (SR) is a nonlinear cooperative e¨ect in which largescale stochastic (random) ¯uctuations (e.g., noise) are entrained by an independent, often periodic, weak ¯uctuation, or signal, with the result that the weak signal ¯uctuations are ampli®ed, with sometimes spectacular results. In other words, the noise displays the same periodic behavior as the signal, even though the signal is very weak, and even though the noise has no periodic behavior of its own. Chapters 21 and 22 discuss many situations of interest to cognitive scientists in which SR arises naturally, among them sensation and perception in general, detection of prey and predators in particular, and the functioning of neurons and groups of neurons, including synchronization of neural oscillations and other exotic phenomena of dynamical computational neuroscience. (Thus, if you need to be motivated to dig into the meat of stochastic resonance, skim those later chapters ®rst, then read this chapter, then go back and read chapters 21 and 22 again, this time with enhanced understanding.) Stochastic resonance was originally proposed as a mechanism whereby large ¯uctuations in the global climate of the earth (i.e., ice ages) could be explained by statistical resonance of the earth's temperature ¯uctuations with small regular ¯uctuations in the amount of solar energy reaching the earth's surface caused by perturbations of the earth's orbit around the

176

Chapter 20

Figure 20.1 (A) Double-well potential; (B) e¨ect of sine wave on probability of movement by a ``kick'' from well to well. (From Gammaitoni et al. 1998. ( 1998 by The American Physical Society.)

sun. The mystery was how such small ¯uctuations in what is called ``insolation'' could cause large ¯uctuations such as the ice ages, during which earth's average temperature was many degrees below normal. To explain this e¨ect, the earth's climate was modeled by a simple physical system called a ``double-well potential'' (®gure 20.1A), in which noiseinduced hopping between the two stable states, ice age and non±ice age, was synchronized with the weak periodic forcing caused by variations in the earth's orbit around the sun (Benzi, Sutera, and Vulpiani 1987). To explain the ice ages, the noise that induced the state changes was taken to be the yearly ¯uctuations in amount of solar radiation, which are not only

Stochastic Resonance

177

random but also quite large and fast compared to the very small, slow, periodic ¯uctuations in solar radiation caused by the changes in the earth's orbit around the sun. Stochastic resonance in the double-well potential system has been very well analyzed. Picture a tiny particle sitting in one of the two depressions (the stable states or ``potential wells'') in the diagram of ®gure 20.1A. If it were not acted on by any force, it would just sit there. However, suppose there is noise (random energy inputs to the system), and every once in a while the particle gets a ``kick'' (perhaps it is hit by a stray photon of light) that causes it to move toward the hump in the middle (the potential barrier). If the kick is large enough, the particle will jump over the hump, out of the well it is in, and into the other one. Clearly, the average time between such well-to-well jumps will depend on how often the particle is kicked by an amount equal to or greater than required to get over the hump. The average waiting time …Tw † is the inverse of the average rate …rK † at which such jumps occur, called ``Kramer's rate'' after the physicist H. A. Kramers (1940), who derived it, and written rK ˆ c
e ÿDV =s ;

…20:1†

where c is a constant that depends on the particular characteristics of the physical system, DV is the height of the potential barrier (see ®g. 20.1A), and e is the base of the natural logarithms …e ˆ 2:7183 . . .†. Thus the average waiting time can be written Tw ˆ

1 1 e DV =s : ˆ ˆ ÿDV =s c rK c
e

…20:2†

Clearly, for a particular double-well system characterized by a potential barrier DV , the average waiting time depends inversely on s, the standard deviation of the statistical distribution of the noise (e.g., the standard deviation of a normal, bell-shaped, distribution): the larger the value of s, the smaller the value of DV =s, thus the smaller the value of e DV =s and the smaller the average waiting time. Now imagine we have a system in which the particle is ``forced'' by another, periodic energy input, say a sin…ot†, where o is some frequency and t is time and a is the (very small) amplitude of the sine wave. This input varies smoothly with time, rising and falling around a mean value. The e¨ect of this varying energy input on the double-well system can be imagined pictorially as causing the double-well system to tilt back and forth, as in ®gure 20.1B, not by enough to cause the particle to jump into

178

Chapter 20

the other well, but just by a little bit. At each instant of time, then, the particle will have some energy a sin…ot† added to it plus another, random kick by the noise. If the total of the two energy inputs is enough to exceed the barrier, then the jump will happen. E¨ectively, the addition of the periodic forcing the a sin…ot† lowers the barrier slightly in one direction or the other, meaning that the kick required from the noise is slightly smaller in the lowered direction (and larger in the other). Now, if the frequency of the tilting is about the same as the rate at which the noiseinduced transitions would ordinarily happen (Kramer's rate, rK ), then the two are said to be in ``stochastic resonance.'' In other words, if the average waiting time between the noise-induced transitions between wells is roughly equal to one-half the period of the forcing, then the noise-induced hopping will be highly likely to happen at just the same time that the bias in one direction or the other is greatest. This hopping is stochastic resonance because a jump does not happen every time; it is only highly likely to happen at certain times, and less likely to happen at other times in between. In fact, the likelihood of a jump at any moment of time exactly mirrors the ups and downs of the sine wave. For this to happen, the noise amplitude, s, must be just rightÐthis is called ``tuning the noise.'' This requirement for tuning the noise is critical to the demonstration of stochastic resonance; the system is resonant with the forcing only for a certain range of s, below which the noise-induced transitions are too rare for the forcing to be visible, and above which the noise-induced transitions are too common and the periodic component is buried in the noise. Stochastic resonance has been demonstrated in many bistable physical systems (those having two stable states), in excitable systems (which have only one stable state and a threshold to an unstable excited state), and in threshold detectors (which have no dynamics but rather ®re a pulse each time an input exceeds a threshold). Gammaitoni et al. (1998) have provided an extensive and detailed review of the many manifestations of SR in these contexts. Because all examples of SR are characterized by maximum sensitivity to a weak perturbation at an optimal level of noise, the notion of stochastic resonance has been generalized to one of information transmission from the weak perturbation through a threshold-dominated system (e.g., Moss, Pierson, and O'Gorman 1994). Because this ``informational'' stochastic resonance is fundamentally a threshold phenomenon, its essential characteristics can be revealed by a study of threshold detectors, such as the one I now describe. (Experimental studies of stochastic resonance in excitable systems such as neurons and people are discussed in chapters 21 and 22.)

Stochastic Resonance

20.2

179

Stochastic Resonance in a Threshold Detector

Consider a detector that ignores any input below a threshold value, a > 0, and that responds with a ``pulse'' (say an output of 1) whenever the input is greater than the threshold. Let there be a weak signal s…t† that is a function of time and that takes values less than a at all times. For simplicity, we will take a ˆ 1 and s…t† < 1, that is, the signal is always below threshold so that, without noise, the detector would not respond to it at all. Let ti ˆ i=n, i ˆ 1 . . . n, be equally spaced time points in a 1 sec sampling interval, where n is the number of samples. Noise is represented by independent random variables e i with common continuous distribution function F. Remember from chapter 14 that a distribution function is the integral of a probability density function considered as a function of the upper limit of integration, xu . Although F can be any distribution function we will assume that it is the distribution function of the Gaussian density function that has mean 0 and standard deviation s. Thus … xu 1 2 2 p
e ÿx =2s dx: …20:3† F …xu † ˆ ÿy s 2p At each time point the input to the detector is the noisy signal s…ti † ‡ e i . Whenever s…ti † ‡ e i > a the detector ®res. I will call these instances ``exceedances,'' indicating that the noisy signal has exceeded the threshold, and they will be represented by the number ``1.'' Thus the output of the detector, Xi , will be 1 if and only if there is an exceedance, and 0 otherwise. According to probability theory (see Greenwood, Ward, and Wefelmeyer 1999), the probability of an exceedance in this situation, p…ti †, is p…ti † ˆ 1 ÿ F ‰a ÿ s…ti †Š:

…20:4†

This situation is illustrated in ®gure 20.2, where the signal is a sine wave and s ˆ 1, and the y-axis represents the amplitude of the noisy signal at each time point. Notice the line at ``1'' on the y-axis: any dot above that line is an exceedance. The number of exceedances per unit time over short segments of time where the signal is roughly the same depends on the distance of the signal from the threshold over that time interval. Moreover, it should also be clear that noise with a smaller standard deviation will lead to less vertical spread of the noisy signal and to fewer exceedances everywhere, whereas noise with a larger standard deviation will have the opposite e¨ect. Thus, for suitable s, the signal is well represented by the time density of exceedances. Figure 20.3 shows the exceedances in

180

Figure 20.2 Noisy sine wave signal.

Figure 20.3 Exceedances from part of signal in ®gure 20.2.

Chapter 20

Stochastic Resonance

181

Figure 20.4 Typical stochastic resonance curve.

part of the signal from ®gure 20.2. It is clear that the time density of these varies regularly, although noisily, with the distance of the signal from the threshold of ``1.'' It is easily possible to determine which part of the sine wave gave rise to this set of exceedencesÐit is the part from the trough of one cycle to the peak of the next. To obtain an expression for an estimate of the signal at any given time point, s^h …ti †, we insert an estimate of p…ti † into equation 20.4 and solve for the estimate of s…ti †, putting the little hats on p…ti † and s…ti † to indicate that these are estimates. The result is s^h …ti † ˆ a ÿ F ÿ1 ‰1 ÿ p^h …ti †Š;

…20:5†

where the h subscript appears because of the way we estimate p…ti †, averaging over a local region of size 2h around each time point (MuÈller 2000). Providing we know what the distribution function of the noise is, and that it has an inverse, we can estimate the signal at each time point by plugging the value of a and F ÿ1 ‰1 ÿ p^h …ti †Š into equation 20.5 and solving. The mean squared error of estimate for the entire signal, MSE…h; t† ˆ E…^ sh …t† ÿ s…t†† 2 , can be calculated approximately for large n for a known signal, or even for various ranges of signals (see MuÈller 2000). In fact, the ``wiggliness'' of the signal (its derivatives) and the standard deviation of the noise distribution both a¨ect the mean squared error for a given sampling rate and threshold. If we assume that the noise is normally distributed, the threshold is 1, and the ®rst and second derivatives of the

182

Chapter 20

signal are between ÿ2 and 2, and then calculate the worst-case approximate mean squared error for a given s, we obtain the values plotted in ®gure 20.4. As it turns out, this is a standard stochastic resonance curve, demonstrating the ``tuning'' of the noise. The y-axis of the graph is proportional to the amount of information recovered about the signal from the exceedances alone, rising steeply from near 0 at low values of s to a peak at somewhat higher values and then gradually falling as s increases further. The peak is at the value of s at which the amount of noise is ``just right'' to reveal the subthreshold signal's ¯uctuations in the exceedances without swamping them. Curves like that of ®gure 20.4 are diagnostic of the informational type of stochastic resonance and similar curves appear in physical stochastic resonance. They have been produced for many types of signals and many types of nonlinear systems, including signals that have no periodic components (random signals) and physical systems that have no strict thresholds (see Gammaitoni et al. 1998 for a review). The above-mentioned method of estimating the signal from the exceedances has been extended to the case of unknown signals with good results (MuÈller and Ward 2000). The technique is complicated, and works better for relatively smooth signals than for the more wiggly ones, but it gives good signal recovery for a wide range of subthreshold signals, and it is based on well-understood statistical procedures (see also Greenwood et al. 1999). Because we can now calculate the expected information recovery whenever a known subthreshold signal and its associated noise are to be estimated, we can know in advance whether a given signal can be e½ciently recovered. These techniques should also be useful in modeling various systems that must contend with weak signals buried in noise, such as those discussed in chapter 21.

Chapter 21 Stochastic Resonance and Perception

Broadband stochastic resonance is manifest in the peripheral layers of neural processing in a simple sensory system, and . . . it plays a role over a wide range of biologically relevant stimulus parameters. ÐJacob E. Levin and John P. Miller, ``Broadband Neural Encoding in the Cricket Cercal Sensory System Enhanced by Stochastic Resonance''

21.1

Detection of Weak Signals by Animals

The watery world of a rippling brook is an inspiration to poets, but a very dangerous place for a cray®sh. It must swim there to obtain its own food, but much larger and faster creatures prowl the same waters, hoping to make a tasty meal of the hapless cray®sh. Although the rhythmical movements made by a swimming ®sh disturb the water ahead of a predatory ®sh, the ®sh often attacks so fast that the information about its imminent arrival is not detected by its prey until too late. In this situation, an early warning system has clear survival advantages. If the faint signals of a predator's attack could be detected early enough, the prey might have enough time to take appropriate action and avoid being eaten. In the case of our cray®sh, there are sensory cells that are specialized to detect weak, coherent water motions such as those produced by a swimming predatory ®sh. The tiny hairs protruding from these cells bend in response to water motion, and this bending generates electrochemical signals, which are transmitted to the ganglia (interconnected groups of interneurons) that control escape movements. The hair cells react best to periodic bending of their hairs by water motion in just the frequency range produced by swimming predatory ®sh. The perennial problem for the cray®sh is, can this sensory system detect the predatory ®sh in time for the cray®sh to escape being eaten? A rippling brook, or any other body of water, is a noisy placeÐit is constantly moving under the in¯uence of so many factors that it is as if

184

Chapter 21

the creatures in it are immersed in a bath of white noise. This seems like the ideal place to exploit stochastic resonance. Perhaps the noisy water acts to amplify the weak signals of the cray®sh's predators, allowing the cray®sh to detect them earlier than it might otherwise with the apparatus at hand. In an attempt to simulate the actual watery environment of the cray®sh, Douglass et al. (1993) generated weak, subthreshold, periodic (sine-wave) water motions like those of a swimming ®sh at a distance, and then added ``white-water noise'' (random water motions like those encountered on a white-water rafting trip) at various amplitudes to create noisy signals like those pictured in ®gure 20.2. In this noisy water, they immersed an excised piece of cray®sh tail, which contained the required sensory hairs, as well as the nerve and ganglia connected to them. Recording the responses of single neurons in the nerve and the ganglia to various water motions, they compared the output of the cray®sh neurons to the noisy weak periodic input signal by calculating the power spectrum (from a fast Fourier transform, or FFT; see chap. 14) of the time series of action potentials emitted by the neuron in response to the noisy signal. Before taking up Douglass et al.'s results, let us brie¯y consider the common technique of power spectrum analysis in stochastic resonance situations. To begin with, if we calculated the power spectrum of the noisy input signal itself, the result would look something like ®gure 21.1, showing a ``spike'' of power at the frequency of the weak periodic signal (here a nominal 1 Hz; ignore the second, smaller spike at 2 Hz, the second harmonic of the signal), and a lower, ¯at, background representing the power spectrum of the noise. The ratio of the height of the spike to the

Figure 21.1 Power spectrum of a noisy periodic signal.

Stochastic Resonance and Perception

185

height of the noise is called the ``signal-to-noise ratio'' (SNR), and can be expressed as a di¨erence in decibels (dB) between signal and noise. Because the y-axis in ®gure 21.1 is in decibels, the SNR of the spike at 1 Hz is just 0 ÿ …ÿ27† or about 27 dB. For the noisy input signal, the SNR would decrease as the noise increased because the background level in ®gure 21.1 would rise, but the signal level would stay the same (at 0). In the time series of action potentials from the sensory neurons exposed to the water motions, the SNR at the signal frequency indicates the extent to which the variation in the neuron's output of action potentials mirrors the ups and downs of the subthreshold periodic signal. A typical power spectrum of such a time series would also resemble ®gure 21.1, except that often the second, smaller spike would also be present. If stochastic resonance were present, a graph of SNR versus noise level for the sensory neurons should resemble ®gure 20.4, at ®rst increasing and then gradually decreasing as the noise level increases. Notice that this is quite di¨erent from what happens with the SNR of the input signal itself, which steadily decreases with increasing noise. It is this paradox that makes stochastic resonance so intriguing. Turning to Douglass et al.'s results (1993), let us look at the SNR for one sensory neuron in their experiment, shown by the solid squares in ®gure 21.2. Each square represents a measurement of the SNR for

Figure 21.2 Stochastic resonance in the cray®sh mechanoreceptor (From Wiesenfeld and Moss 1995. Reprinted with permission from Nature. ( 1995, Macmillan Magazines, Limited.)

186

Chapter 21

responses of the neuron to the noisy signal for a di¨erent amplitude of water noise. Stochastic resonance is clearly present, for the SNR of the neuron's response reaches a peak at an intermediate level of water noise. The open squares in the ®gure represent the SNR values for a computer simulation of a model neuron (see chap. 22; Moss et al. 1993) and the solid line a theoretical curve that characterizes many systems displaying stochastic resonance. The actual neuron responses are highly similar to those of the simulated neuron except at low noise levels. Douglass et al. speculated that this is because the actual neuron also has internal noise that may increase its ability to follow a subthreshold signal even in the absence of external water noise. Other animals also seem to use stochastic resonance in a similar way to avoid predators or to capture prey. For example, it is thought that crickets avoid wasp predators at least partially through using stochastic resonance to detect very weak air currents generated by their attackers (Levin and Miller 1996). In this case, through similar experiments, Levin and Miller demonstrated that a combination of random air currents and the faint vibrations made by a predatory wasp's wings can be detected by the cricket's cercal hair cells well before they become detectable in the absence of the ``air noise.'' Moreover, the noise-induced enhancement of coherence of the cricket's sensory response with the subthreshold signal occurred in just the frequency range of air movements generated by the wasp predator, and for just the same faint intensity of air motions where crickets really do escape from wasps. It thus seems plausible that the mechanism is actually in use by these noisy creatures. Stochastic resonance appears in mammalian sensory systems as well. Collins, Imho¨, and Grigg (1996a) excised a piece of a rat's skin, along with its mechanoreceptors and the slowly adapting nerve ®ber that comprised the axons of the receptor neurons, placed it in an appropriate physiological apparatus, and made recordings of individual receptors' action potentials emitted in response to precisely measured touches on the piece of skin. The touches varied in intensity corresponding to a nearthreshold, slowly but randomly varying signal to which the receptors seldom responded. Random ``touch noise'' was also added to the skin stimulus. The correlation between the resulting noisy touch signal and the momentary ®ring frequency of the touch neurons was maximal for an intermediate level of touch noise, again displaying stochastic resonance, in this case, for an aperiodic signal. Such experiments show that animals can use stochastic resonance in their sensory systems to avoid predators (or detect prey) but do they

Stochastic Resonance and Perception

187

actually do so? The ®rst experiment to show that an animal does use stochastic resonance to its advantage was reported by Russell, Wilkens, and Moss (1999), who put juvenile paddle®sh into a stream of ¯owing, electrically noisy, water, and introduced daphnia, tiny plankton eaten by the paddle®sh, into the stream. In nature, a juvenile paddle®sh uses an array of electrosensors on its long snout (rostrum) to detect daphnia, which it then pursues and eats. Russell, Wilkens, and Moss showed that the paddle®sh were able to capture more daphnia presented farther from the rostrum in moderately electrically noisy water than in electrically quieter or noisier water, indicating that stochastic resonance was enhancing detectability of the faint electrical signals generated by the daphnia. Greenwood et al. (2000) then showed that the relationship between the probability of a paddle®sh striking at a daphnia and the distance of the daphnia from the rostrum, in the presence of the optimum amount of noise, can be explained by a simple theory of how stochastic resonance enhances information extraction from subthreshold signals. An example more relevant to humans concerns the isolated sciatic nerve of the toad Xenopus laevis, whose responses are used to predict those of the human auditory nerve to vowel sounds coded by cochlear implants. Surgically implanted in the cochleas (the inner ear structures that transform sound waves into electrical impulses) of profoundly deaf people, cochlear implants electrically stimulate the auditory nerve in response to external sounds. Morse and Evans (1996) found that, although standard analog cochlear implants did not do well at stimulating a useful pattern of nerve activity, adding noise to the speech stimulus enhanced the encoding of distinguishing acoustic features of vowels in the pattern of toad sciatic nerve neural activity. Subsequently, Zeng, Fu, and Morse (2000) showed that adding noise to auditory stimuli did indeed improve the ability of cochlear implant recipients to detect and discriminate the frequency of pure tones. This and other results suggesting stochastic resonance led Morse and Evans (1996), Moss, Chiou-Tan, and Klinke (1996), and Zeng, Fu, and Morse (2000) to propose that noise be added deliberately to the input of cochlear implants to enhance the ability of those receiving them to distinguish speech and other natural sounds. Cordo et al. (1996) demonstrated that stochastic resonance occurs in human sensory systems too, using muscle spindle receptors, which detect stretching of the muscle. They recorded the responses of a single a¨erent ®ber from the wrist and hand extensor muscles of each subject while the wrist was being passively rotated (to stretch the muscles) back and forth in a sinusoidal motion that covered about 2 degrees. They generated

188

Chapter 21

``motion noise'' by randomly vibrating the tendon attached to the extensor muscle, causing the muscle to stretch randomly. The very tiny rotary movement caused very few action potentials in the muscle spindle neuron in the absence of the motion noise and thus a very low SNR (near 0 dB) at the frequency of the weak periodic signal in the power spectrum of the time series of action potentials. When, however, they introduced the motion noise, the SNR increased, with the maximum SNR occurring for a moderate level of motion noise. For even higher levels of motion noise, the SNR gradually decreased, as the noise began to dominate the response of the neuron, indicating stochastic resonance was operating. Cordo et al. argued that internal noise generated by the intrafusal muscle ®bers on which the stretch receptors lie provides the stretch receptors with the optimum amount of motor noise so that near-threshold movements can be detected and signaled by the stretch receptors, enhancing motor control. 21.2

Stochastic Resonance in Human Perception

Stochastic resonance also operates at the level of the whole organism in human perception. Perhaps the most dramatic example was provided by Simonotto et al. (1997). Although not used by them, ®gure 21.3 illustrates what they found. It contains three pictures of Big Ben, the famous London clock tower. To produce these, Simonotto et al. ®rst digitized a standard photograph of Big Ben on a 256-level gray scale with a spatial resolution of 256 by 256 pixels. (A pixel is a picture element, usually a tiny square, which has the same color, or gray-scale value, everywhere, but can be di¨erent from its neighbors.) Then, to the gray scale value in every pixel,

Figure 21.3 Big Ben is seen best in optimal noise. (From Simonotto et al. 1997. ( 1997 by The American Physical Society.)

Stochastic Resonance and Perception

189

initially I < 30, was added noise, which consisted of a random number, x, drawn from a normal distribution with zero mean and standard deviation s. The resulting noisy gray scale values, I ‡ x, were then replaced by white if I ‡ x < D and by black if I ‡ x b D. In the black-and-white pictures shown in ®gure 21.3, s ˆ 10, 90, and 300, respectively, from left to right. As you can see, Big Ben is clearest when s ˆ 90, an intermediate level of gray-scale noise, displaying stochastic resonance in the picture-making process. In this case, the threshold was D ˆ 30, a level that was clearly too high to yield a recognizable picture when the noise level was low …s ˆ 10†. At an intermediate noise level, however, the subthreshold picture is clearly apparent, whereas at a higher noise level it is beginning to be masked by the noise and it is not as clear. In their experiment, Simonotto et al. (1997) created patterns of stripes by varying gray-levels sinusoidally across the face of a computer monitor, to which they added various amounts of noise that varied over time (dynamic visual noise; see Uttal 1969). They turned each gray-level pattern into a black-and-white pattern by applying a threshold, much as in ®gure 21.3. Each resulting dynamically noisy striped pattern was created in seven di¨erent degrees of contrast. The contrast of a pattern is the ratio of the luminance of its lightest part to that of its darkest part. Using grayscale values and calling black ``1,'' when the lightest part has a gray-scale value of 25 and the darkest part is black, the contrast is 25/1 or 25. The higher the gray-scale value of the white part, the greater the contrast. Lower contrast patterns are harder to see. Simonotto et al. asked their subjects to choose the lowest contrast noisy pattern for which they could just distinguish a particular ®ne detail (very thin stripe). This ``contrast threshold,'' which represents the best resolving power of the visual system for these stimuli, varied with the intensity of the noise in the way predicted by stochastic resonance. That is, it was lowest (resolving power was highest) for an intermediate level of noise. Indeed, Simonotto et al. found that a simple model of threshold stochastic resonance ®tted their data extremely well: Ath ˆ Ks exp‰D 2 2s 2 Š

…21:1†

According to this model, subjects detect the signal power, Ath , rather than the signal-to-noise ratio, where Ath corresponds to the contrast threshold mentioned above. Clearly Ath depends on s, which represents the amount of noise and on K, which represents the extent to which di¨erent subjects were sensitive to ®ne details in a noisy pattern (ignore D, which has a technical meaning). Small K values mean low thresholds and high sensi-

190

Chapter 21

tivity. Of course, s varied across patterns to give the stochastic resonance curve. That the model of stochastic resonance used was based on the power spectrum of a time series of identical pulses similar to the action potentials of neurons suggests that the visual system might perform similar computations when we look at noisy images. Another example of perceptual stochastic resonance comes from a study of tactile sensation. As in their work with rat skin, Collins, Imho¨, and Grigg (1996b) touched their human subjects' skins with a probe. In this experiment, however, the skin remained attached to the subjects, who were only asked to say whether they had detected a subthreshold touch signal added to ``touch noise'' or only the noise itself. The touch signal was a constant touch applied for a quarter of a second at a force that was just subthreshold in the absence of noise. Collins, Imho¨, and Grigg (1996b) found that, as input noise level increased, performance rose from chance (at zero noise) to signi®cantly above chance at moderate noise levels, and then declined to below chance again. Thus people seem to be able to exploit stochastic resonance in the touch perception system to improve their detection of faint touches. In addition to their studies with cochlear implant recipients, Zeng, Fu, and Morse (2000) also showed that adding subthreshold noise to subthreshold tones allowed normal-hearing subjects to detect them. In my own laboratory, we are studying the use of stochastic resonance to improve normal-hearing subjects' detection of faint sinusoidal modulations of clearly audible sounds, reasoning that such abilities are most likely to have had evolutionary signi®cance. In our experiments, human subjects are asked to tell which one of two roughly 70 Hz sounds is beating at 3 Hz. Various levels of noise are added to both sounds so that the only di¨erence between them is the presence of the beats in one sound. Thus far, we have evidence that detection of the beats, measured with a bias-resistant measure, is best at nonzero levels of added noise, although the e¨ect is small and very di½cult to measure (Ward, Moss, Desai, and Rootman 2001). Similarly, when human subjects try to discriminate which of two patches on a computer monitor contains a square-wave grating (or a sinusoidal grating), they perform best at nonzero noise levels (Ward, Moss, Desai, and Tata 2001). Again, this e¨ect is small and di½cult to measure, but with sophisticated psychophysical techniques we have obtained the data shown in ®gure 21.4 (only square-wave grating detection shown). The mean contrast threshold data clearly show the prototypical stochastic resonance curve. Moreover, a model very similar to that of Simonotto et al. (1997) also ®ts these data very well (solid curve in ®gure 21.4). One

Stochastic Resonance and Perception

191

Figure 21.4 Stochastic resonance in the human visual system.(From Ward, Desai, Tata, and Moss 2001.)

implication of these demonstrations is that stochastic resonance could prove to be useful in ameliorating sensory de®cits produced by aging or damaged receptor systems, in which sensory thresholds are raised, even in the absence of prostheses such as cochlear implants. It is conceivable that by adding the right amount of noise to sensory signals, they could be usefully ampli®ed beyond the raised thresholds, enhancing perception of the previously subthreshold signals. Stochastic resonance in human behavioral performance is somewhat controversial because the standard model used to measure performance in a bias-resistant way, signal detection theory (SDT; e.g., Green and Swets 1966; Macmillan and Creelman 1991), predicts that there should be no such thing as stochastic resonance (Ward 1999; Tougaard 2000). There is no threshold in SDT: the theory assumes that all signals will add an amount of energy to the background noise distribution such that the distribution of signal plus noise can always be discriminated from the distribution of noise alone given a su½cient number of samples from each. In the face of demonstrations of stochastic resonance in human perception, we need a way to understand the situation that both preserves the biasresistant nature of SDT measures and also allows stochastic resonance. Moss, Neiman, and I (Ward 2000; Ward, Neiman, and Moss 2001) have argued that if any element of the signal processing system has a signi®cant

192

Chapter 21

nonlinearity, such as a hard threshold (like that of the detector described in chap. 20), a soft threshold (a relatively steep but not discontinuous increase in the probability of detection as a function of input signal strength) or even a su½ciently nonlinear transfer function (an exponential will do), then stochastic resonance will result, even in a bias-resistant measure of performance such as that o¨ered by signal detection theory. Demonstrations of this e¨ect in two speci®c situations can be derived from the work of Bezrukov and Vodyanoy 1997 and in Stemmler 1996. Thus there are both demonstrations of stochastic resonance in human performance and the beginning of a theoretical understanding about how weak signals can be made detectable by adding optimum amounts of noise, creating an intriguing oxymoron: ``useful noise.''

Chapter 22 Stochastic Resonance in the Brain

As neurons in higher centers of the brain need to maintain a high signal-to-noise ratio as well as peripheral ones, it is plausible to presume that stochastic resonance is a general principle of biological information processing. ÐH. E. Plesser and S. Tanaka, ``Stochastic Resonance in a Model Neuron with Reset''

22.1

Stochastic Resonance in Neurons

Although, under the appropriate conditions, stochastic resonance may be exhibited by any level-crossing detector, it may play a particularly profound role in excitable systems such as cortical interneurons. A cortical interneuron produces an action potential (or spike) when the membrane electrical potential at the ``trigger zone'' of the neuron (where the axon extends from the soma) depolarizes to just above the resting potential of ÿ70 mV, namely, at about ÿ60 to ÿ50 mV. The membrane potential, the voltage across the membrane, from the outside to the inside of the neuron, is negative when the neuron is ``resting'' (no inputs) because of a dynamic equilibrium of ion ¯ows across the membrane and the presence of large, positively charged molecules inside the neuron. Each action potential triggered by a change in membrane potential at the trigger zone lasts about 1 to 2 msec, and the neuron's threshold for ®ring is higher than usual for an additional 1 to 2 msec after the end of the action potential. Thus, under optimal conditions, the neuron can ®re at most only about every 2 msec, making the maximum ®ring rate about 500 Hz. Changes in membrane electrical potential relevant to action potentials are produced by excitatory (depolarizingÐmaking membrane potential more positive) and inhibitory (hyperpolarizingÐmaking it more negative) postsynaptic potentials (EPSPs and IPSPs, respectively) generated when excitatory or inhibitory transmitter substance is released across a dendritic synapse

194

Chapter 22

from the axonal endings of another neuron. EPSPs and IPSPs can directly a¨ect the generation of an action potential in a resting neuron or can simply modulate a neuron's spontaneous rate of ®ring (see Eccles 1964 for details on postsynaptic potentials and neuron ®ring). Each release of transmitter substance causes a postsynaptic potential that quickly reaches a peak and then decays exponentially in time. Such membrane potential changes are largest near the receptor site and propagate to and across the soma and down the axon, decaying exponentially with distance from the source. EPSPs and IPSPs are usually thought to summate algebraically in their e¨ect on the total membrane potential and across space and time. Thus inputs to the neuron that generate postsynaptic potentials while the e¨ects of earlier ones are still present will add to those e¨ects, possibly increasing or decreasing the likelihood of exceeding the neuron's threshold for an action potential. Koch (1997) has observed that even more complicated interactions can take place between postsynaptic potentials at di¨erent sites on the neuron, including multiplicative interactions. Hodgkin and Huxley (1952) were the ®rst to propose a mathematical model of the events surrounding the generation of, and recovery from, action potentials in neurons. Their model consists of four di¨erential equations that faithfully reproduce the series of action potentials seen in real neurons, while also describing in some detail how inputs from other neurons and various membrane currents a¨ect them. A simpler model of neural ®ring (involving only two di¨erential equations) that preserves most of these features was suggested by Fitzhugh (1961) and by Nagumo, Arimoto, and Yoshizawa (1962). (This model is studied further in chapter 31 in the context of relaxation oscillators.) Finally, an even simpler model, called the ``integrate-and-®re'' model, describes only the ability of the neuron to integrate various subthreshold inputs over time and to ®re a series of action potentials. A di¨erential equation for this last, simplest model is dv ˆ ÿav ‡ a ÿ d ‡ S…t† ‡ x…t†; dt

…22:1†

where v is the voltage across the membrane (membrane potential), a and d are constants, S…t† is a subthreshold slowly varying but random signal representing the inputs of various other neurons to the neuron, and x…t† is normally distributed noise that at each time is slightly correlated with the values at previous times (e.g., Collins et al. 1996; Plesser and Tanaka 1997). The voltage, v, is reset to its starting value (usually 0) every time it exceeds the threshold and an action potential is generated. Another way

Stochastic Resonance in the Brain

195

Figure 22.1 Membrane voltage ¯uctuations for an integrate-and-®re model neuron.

to write this model is as a time series, vt‡1 ˆ vt ÿ avt ‡ a ÿ d ‡ St ‡ xt ;

…22:2†

where the membrane potential at time t ‡ 1, vt‡1 , consists of the value at time t, vt , plus the change over a small time interval given by equation 22.1, and typical values for the constants are a ˆ 0:5 and d ˆ 0:01. Figure 22.1 shows how the membrane voltage in this model ¯uctuates over time (200 time periods are shown from a much longer run), given these values of the constants and a particular series of values for St and xt . The threshold for ®ring an action potential in the integrate-and-®re model is set as a constant, for example v ˆ 1. Considering this threshold for the series of voltages displayed in ®gure 22.1, you can see that a series of action potentials resulted whenever the voltage exceeded 1, separated by periods of time during which no action potentials occurred, much like the picture in ®gure 20.3. An interesting aspect of this integrate-and-®re model of a neuron is that the signal, S…t†, can be adjusted to be small enough that in the absence of the noise, x…t†, the voltage will never cross the threshold and the neuron will never ®re. This is just the situation in stochastic resonance. Collins et al. (1996) showed that the mean ®ring rate of the integrate-and-®re model neuron (similar to the mean number of spikes in a short interval around the time of interest, for each time point in the series) tracked the variations in such a subthreshold signal best, and reasonably well at that, for an intermediate value of the noise intensity

196

Chapter 22

Figure 22.2 Signal and spikes for the same time periods as in ®gure 22.1.

(standard deviation of the distribution from which the noise was drawn). An example of such tracking, shown in ®gure 22.2, illustrates that when the (subthreshold) signal was relatively high, toward the left side of the graph, the frequency with which spikes were generated by the model neuron was much higher than when the signal was relatively low, toward the right side of the graph. A similar demonstration for a subthreshold periodic signal was accomplished by Plesser and Tanaka (1997). Thus the model integrate-and-®re neuron exhibits stochastic resonance; the same is true for the more complicated model neurons described earlier (e.g., Collins et al. 1996; Longtin 1993, 1997). Thus, too, real cortical interneurons may also exhibit stochastic resonance because these models are quite good descriptions of their dynamics. Indeed, researchers such as Plesser and Tanaka (1997) in the epigraph above have speculated that stochastic resonance is a general principle of biological information processing. 22.2

Neural Networks

Because each cortical interneuron receives inputs of transmitter substance from many other neurons (about 1,000 on average) via dendritic and somatic synapses with their axons, each is immersed in a complex envi-

Stochastic Resonance in the Brain

197

ronment of external signals. From the point of view of a single neuron sending a coherent signal to another neuron, that signal is in turn immersed in the noise generated by all of the other connecting neurons and other surrounding support cells and the intercellular medium, as well as by metabolic processes of the recipient neuron. Because all of the postsynaptic potentials summate (or even multiply) in their e¨ect on action potential generation, and because a single action potential input from another neuron typically has a subthreshold e¨ect on the total membrane potential (e.g., Thomson et al. 1996), the situation is exactly right for stochastic resonance to amplify a subthreshold signal above the action potential threshold. If all of the other inputs to a neuron are incoherent with each other, then they constitute noise. In such a context, a coherent input signal, even if subthreshold, could be ampli®ed above the threshold by that noise and drive the output of the target neuron, in much the same way described for the integrate-and-®re model neuron. This scenario was studied by Mato (1998) using such a model neuron, which received two input spike trains: one was periodic and subthreshold, and the other random with a given variability. Because the random input represented the unsynchronized inputs of other neurons, its variability represented the number of other, unsynchronized neurons actively stimulating the target neuron. The target neuron exhibited typical stochastic resonance behavior in that it tracked the periodic input signal best at an intermediate level of random input noise. Moreover, this ®nding demonstrated that the more usual scenario, in which many inputting neurons synchronize their inputs so as to deliver enough action potentials to the target neuron within a small enough time window to excite action potentials postsynaptically, is not the only way in which neurons can be driven by input signals. In the stochastic resonance scenario, one or a few neurons could drive the output of another neuron as long as they produced a synchronized input signal. The output of each neuron could also represent a blend of several di¨erent, subthreshold, input signals, each incoherent with the others, but all producing a complex, additive periodicity in the output of the target neuron. This process would happen best when just the right number of incoherent inputs (optimum noise) was being received. As I mentioned earlier, a neuron integrates inputs over time as well as over space. The time constant, t, over which such integration occurs varies from about 100 msec for inputs to sensory neurons, to 40±60 msec for inputs to the soma of pyramidal interneurons, to as short as 10 msec for inputs to the apical tufts of dendrites of pyramidal neurons in the prefrontal cortex (Seamans, Gorelova, and Yang 1997). Larger values of

198

Chapter 22

t make neurons act as integrators, whereas smaller values, such as those at the apical tufts, make them act as coincidence detectors (KoÈnig, Engel, and Singer 1996). EPSPs and IPSPs, as well as action potentials, are produced by ¯ows of ions across the cell membrane of the neuron. In many cases, a cascade of chemical reactions (second messenger system) is begun by reception of transmitter substance at the cell membrane. It is the time courses of such processes that set the value of the time constant t. The electrochemical state of the neuron at any time is a complex balance of electrochemical forces in¯uenced by synaptic and internal events that have occurred over a time window of width t plus the constant di¨usion of ions caused by electrical and concentration gradients and active pumps. The instantaneous rate of ®ring (1,000/T, where T is the interval between two successive action potentials in msec) of the neuron re¯ects all of these forces, and thus can be modeled as depending on inputs over a moving window of length t. Temporal integration in a neuron is represented in all of the model neurons I have mentioned, especially of course in the integrate-and-®re neuron. It has been shown that temporal integration in the form of kernel estimation (see chap. 20) allows even a simple threshold detector to produce outputs that closely follow a nonstationary subthreshold signal, at least at frequencies below the Nyquist cuto¨, which is 2/t in this case (MuÈller and Ward 2000). Thus, for example, a particular neuron might produce spike trains that gradually increase in time density (number of

Figure 22.3 Stochastic resonance in the rat hippocampus. (From Gluckman et al. 1996. ( 1996 by The American Physical Society.)

Stochastic Resonance in the Brain

199

spikes per unit time) as the relatively small number of neurons synapsing with it, and responding in synchrony to an external stimulus, increases. Or its output could simply re¯ect an overall increase in the ®ring rate of the synchronized, driving neurons, which could in turn re¯ect the increasing intensity of the faint noises made by an approaching predator. The growing body of theoretical work on model neurons has made it highly desirable to study stochastic resonance experimentally in neurons in the brain. To address the di½culty in delivering di¨erent levels of signal and noise to these neurons, Gluckman et al. (1996) used a technique adapted from their study of epileptic-like activity in slices of rat brain: they imposed an electrical ®eld on an entire network of neurons in a slice from the rat hippocampus. They did this by placing the slice between two electrode plates, which produced an electrical ®eld in the ¯uid bathing the neurons. The ®eld induced ionic currents both inside and outside the neurons, and e¨ectively shifted the spike threshold. The ®eld was varied sinusoidally in a way that produced no action potentials, and then noise was added to this periodic variation in various amounts. In contrast to previous studies, however, the behavior of individual neurons was not studied in this experiment. Instead, the synchronized bursting behavior of the entire network of neurons, which occurs spontaneously and is an emergent property of the network, was recorded as the electrical ®eld varied. Burst ®ring occurs when neurons ®re several spikes spaced closely in time (a ``burst''), with relatively longer time periods between the bursts. Stochastic resonance has been demonstrated in models of such neurons, even without external forcing (e.g., Longtin 1997) but never before in real bursting neurons. As the top half of ®gure 22.3 shows for the case of the hippocampal slices, when the electrical ®eld contained noise but no signal (Asin ˆ 0, Anoise ˆ 10), synchronized bursts were also random. The bursts are indicated both by the ®eld component graph at the top of each segment of the ®gure and by the graph just above the signal graph (the latter is just a straight line in the top half of the ®gure). But when there was a subthreshold signal, shown in the bottom half of the ®gure where Asin ˆ 3:75 < Athreshold ˆ 7 and Anoise ˆ 10, the bursts tended to occur at the peaks of the signal (i.e., phase-locked to the signal). This tendency was most pronounced for intermediate levels of the electrical noise: stochastic resonance again. Thus it seems that groups of neurons act like individual neurons in displaying stochastic resonance. It seems possible that stochastic resonance is a general property of brain activity, although just how extensive a role it plays is not yet known.

This page intentionally left blank

Chapter 23 Chaos

For my part, at least, I am convinced that He doesn't throw dice. ÐAlbert Einstein And then we see unexpected things . . . our imagination is stretched to the utmost, not, as in ®ction, to imagine things which are not really there, but just to comprehend those things which are there. ÐRichard Phillips Feynman, Character of Physical Law

23.1

Chaos Is Not What You Think It Is

At one time, chaos theory seemed like the answer to Einstein's quest for a theory that would at the same time be deterministic and yet give the appearance of randomness (in order to satisfy the requirements of quantum physics). In the late 1980s, it captured the imagination of the media, and of many scientists; as a result, many excellent popularizations have been written about it (e.g., Gleick 1987; Stewart 1989). There are also deep texts if you have the stomach for them (e.g., Devaney 1989; Thompson and Stewart 1986). In this chapter I will merely skim the surface, introducing some of the ideas critical to understanding what occurs in the following few chapters, which are by way of a progress report. Far from being dead, chaos theory is alive and well and being applied fruitfully in many areas, including dynamical cognitive science. As we might have guessed, however, it has proved not to be a panacea, and many matters that were before mysterious remain so, despite the best e¨orts of many brilliant minds. My desk dictionary (Merriam-Webster 1971, 81) de®nes chaos as ``the confused unorganized state existing before the creation of distinct forms'' or as ``complete disorder.'' And that is indeed part of what chaos theory is about: the surface appearance of disorder, even randomness. But chaos theory is really about nonlinear dynamical systems and their maps, which

202

Chapter 23

are the mathematical functions that determine the behavior of the system. One of the most important goals of the theory of dynamical systems is ``to understand the eventual or asymptotic behavior of an iterative process'' (Devaney 1989, 17). The dynamical system could be speci®ed by a system of nonlinear di¨erential equations in time, in which case understanding the system involves solving the equations, a very di½cult, often impossible undertaking. (For cases where solution is possible, and for techniques to use when solution is too di½cult, see, for example, Devaney 1989; Morrison 1991; see also chaps. 8±10 and 31.) Alternatively, and what we are most concerned with in this chapter, the system could be speci®ed as the iteration of a discrete function, where we want to know what happens to the value of the function as we continue to iterate it. Because iteration is not something we are used to doing with functions, I will tell you a bit more about it before we go on into chaos. To iterate a function f …x† means to calculate the value of the function for some starting value, x, and then to put that value, called f 1 …x†, back into the function in place of the initial x value, so we have f 2 …x† ˆ f … f 1 …x††, f 3 …x† ˆ f … f 2 …x†† ˆ f … f … f 1 …x†††, and so forth. Notice that the superscripts do not have their usual meaning of multiplication here, but rather refer to the number of times a function has been applied to the starting value of the independent variable. Let us look at, for example, what happens to the map f …x† ˆ x 2 if we start with x ˆ 2. (``Map'' is just another word for function because functions map a set of numbers into another set in the sense of connecting pairs of numbers.) Clearly, the value of the function increases without limit. (2 2 ˆ 4, 4 2 ˆ 16, 16 2 ˆ 256, etc.) Indeed, it will do so for any starting value greater than 1 or, after the initial iteration, for any starting value less than ÿ1, in what is called an ``exploding process.'' Nonetheless, this outcome is extremely sensitive to the starting value. If we start with x ˆ 0:1, the process dives toward an asymptote of 0 (0:1 2 ˆ 0:01, 0:01 2 ˆ 0:0001, 0:0001 2 ˆ 0:00000001, etc.), in what is called a ``collapsing process.'' The same outcome will occur for any number between ÿ1 and 1 except for 0. If x ˆ 0, since 0 2 ˆ 0, the system will stay at 0. Similarly, if x ˆ 1, then the system stays right at 1, since 1 2 ˆ 1. Any such points, for which f …x† ˆ x, are called ``®xed points'' of f …x†. Finally, for x ˆ ÿ1, the ®rst iteration of f …x† ˆ x 2 will produce …ÿ1† 2 ˆ 1 and any subsequent iterations will then stay at 1. In this case, the behavior of the function as it is repeatedly iterated is clear and easy to see. If we knew that our dynamical system was described by this function, we would simply have to know the starting value to be able to predict how it would behave, as summarized in table 23.1. Notice that

Chaos

203

Table 23.1 Behavior of Iterated Function f …x† Starting x

2 >2

2 2 ÿ3:72 >ÿ3:72

White noise

A0

A0

p, then the generated series has too many runs. If l ˆ p, then the successive guesses are independent, a random series. The parameter l is easily estimated from the generated series by l^ ˆ a ``heads heads'' / (a ``heads heads'' ‡ a ``heads tails''), where a ``heads heads'' means number of pairs with a ``heads'' followed by a ``heads.'' This model captures both the departures from randomness in subjects' generated series for a wide range of values of p, and also the tendency of subjects to use the response alternatives equally over short segments of the series (``local representativeness,'' one of the heuristics described by Tversky and Kahneman 1981). On the other hand, the model tells us neither why subjects would

258

Chapter 29

choose a particular value of l nor why, when they know they are supposed to make l ˆ p, they usually do not do so. 29.2

Sometimes

In chapter 11, we encountered time estimation data from three subjects who had produced series of 1,000 ``random'' intervals in which the autocorrelation function was null, an indication that these subjects were capable of producing random behavior (although the data were not subjected to more stringent tests of randomness). Three other subjects were unable to produce random time series: their series were well described by ARIMA (0, 1, 2) or ARIMA (1, 1, 1) models. In Budescu 1987, three of eighteen subjects generated at least a few series of coin tosses that had no consistent bias, an instance of the randomness demonstrated by a few subjects in most experiments. Are these subjects di¨erent? Do they have a talent the other ones lack? Or does this random behavior simply occur at random? Not trusting our subjects' answers (for good reasons), psychologists do not usually them to report either on their understanding of the task before them or on the strategy they propose to use to deal with it. We attempt to infer these from their performance or, better, to manipulate them explicitly. In an attempt to test the idea that subjects' nonrandom performance re¯ected imperfect understanding of the concept of randomness (a problem that sometimes a¨ects even mathematicians), Neuringer (1986) conducted a sophisticated computer-based experiment. He asked subjects to press the ``1'' and ``2'' keys on a computer keyboard as randomly as possible and at their own speed. As a baseline, he collected 6,000 responses, in 60 blocks of 100 trials, from each of seven subjects. After subjects had generated these series, he began to give them feedback as to how well their series matched one generated by a random number generator implemented on the same computer. In one experiment, he gave feedback on ®ve di¨erent series descriptors: uncertainty of successive responses (or informationÐsee Garner 1962), uncertainty of responses two trials apart (RNG2), number of runs (ALTS), uncertainty of responses at the same position across 100-trial blocks, and uncertainty of responses at positions two trials apart across 100-trial blocks (XC2). In another, he gave feedback on a group of ten descriptors, the ®rst ®ve and ®ve additional ones that implemented an even more demanding criterion of randomness. He compared each subject's series, both before and after feedback was initiated, to the outputs of the computer random number generator using the

Can People Behave Randomly?

259

Figure 29.3 One subject's data. (Data from Neuringer 1986.)

Kolmogorov-Smirnov test (a nonparametric test of proportions; see Siegel 1956). During the baseline period, all subjects' series di¨ered signi®cantly from the random number generator on most of the descriptors. For example, in the ®rst experiment, all subjects di¨ered from the random number generator on both of the ®rst two descriptors, and four of the seven di¨ered signi®cantly on each of the other three. Interestingly, all of Neuringer's subjects (1986) displayed l > p behavior, to use Budescu's schema (1987), generating too few alternations, the opposite of the usual ®nding. Neuringer speculated that this may have occurred because of the ease of quickly generating strings of the same response on the computer keyboard. Figure 29.3 displays the baseline data for three of the descriptors for one typical subject in the form of the ratio of the subject's score on a descriptor divided by the random number generator's score on that same descriptor: random performance would be indicated by a ratio of 1.0. Each point in the ®gure is an average over twenty 100-digit blocks, (called ``trials'' by Neuringer). Figure 29.3 is also representative of the general ®nding of both experiments: after feedback was initiated, eventually all subjects were able to generate series that passed all of the tests for randomness for at least two successive sets of 60 ``trials.'' For most subjects, this took a long time (remember a ``trial'' refers here to a 100-digit series). The subject shown in ®gure 29.3 required about 300 such ``trials'' to reach consistent randomness; other subjects required over 500 (or

260

Chapter 29

50,000 responses). In the second experiment, with ten descriptors, three subjects required nearly 1,000 such ``trials''; one even required some 1,700 ``trials'' (170,000 responses) after feedback was initiated. Thus, when humans are told exactly what to do, and given feedback on how they are doing, they can even behave randomly, if given enough time to learn how to adjust their behavior to the requirements. Although people must be extensively trained to achieve the most stringent standards of randomness, there is evidence that they do have a reasonable idea of what random behavior is, at least in the sense of relative frequency. As we have noted, small di¨erences are more common than large ones in a random sequence of more than two alternatives. There are similar ``proportions'' in the various subsequences of longer random sequences. Consider a long sequence of ``fair'' coin tosses and subsequences of ten tosses in the longer sequence. There are 2 10 ˆ 1;024 possible such subsequences, and of these, 252 have ®ve heads and ®ve tails, and 210 have either 4 heads and 6 tails or vice versa. Clearly, a 10toss subsequence with about equal numbers of heads and tails is by far the most common type: 67.2% of the possible subsequences are of this type. Kareev (1992) argued that people know that such roughly equally balanced subsequences are most common in a longer sequence, although of course he did not ask them directly. Instead, he reasoned that young children would be less likely to know this fact than college-age adults. Kareev argued that, when asked to produce a random sequence, people try to produce subsequences resembling the most common one in longer sequences; and that they are limited by short-term memory capacity to producing a subsequence of a length consistent with that capacity (between ®ve and seven items for most people). Tune (1964) had proposed a similar explanation for violations of randomness based on short-term memory limitations. Kareev (1992) hypothesized that, because children also have less short-term memory capacity than adults, they would be limited to even shorter subsequences, thus would depart from randomness more often than would adults. He con®rmed this hypothesis for short sequences of ten hypothetical coin tosses; analyzing the alternation rate in sequences of hypothetical coin tosses of various lengths for various probabilities of heads in the sequence, he showed that the most commonly reported alternation rate of about 0.6 is the expected one for a short-term memory capacity of between ®ve and six items. Finally, he showed that the alternation rate observed in his data decreased with increasing age of subject, and was smaller for a task that put more emphasis on each individual trial than on the entire sequence. For the adults doing the trial-by-

Can People Behave Randomly?

261

trial task, it was 0.53, just about the expected 0.5 for a truly random sequence, but still slightly biased in the usual way. The theoretical approach taken in chapter 28 for perception of sequences would seem appropriate for production of sequences, as well. Although it is hardly a well-worked-out theory, associating the contingency estimation processes described in chapter 28 based on long-term and short-term memory, respectively, with the stabilization and tracking processes of Treisman and Williams (1984), goes some way toward making the approach more sophisticated. Treisman and Faulkner (1987) showed that by assuming subjects possess an internal generator of randomly varying activation (internal noise?) that can be categorized like that produced by an external stimulus under the in¯uence of the e¨ects of stabilization and tracking on criterion settings, one can account for all of the previous data and their own new data on random sequence generation. And their theory is very well worked out indeed. An alternative is the theory my colleagues and I proposed in Ward et al. 1988, a di¨erent generalization of SDT with an explicit memory parameter that, when applied to the generation of sequences, produces results roughly similar to those of Treisman and Faulkner (1987). 29.3

Sequential Dependencies and Extrasensory Perception

Many people associate Duke University with the work of J. B. Rhine on parapsychology, although he left there in the early 1960s to start an independent foundation located in Durham, North Carolina, for research into unexplained phenomena, mostly parapsychology. As a graduate student in experimental psychology at Duke, supervised by G. R. Lockhead, I became very aware of the mixed feelings that Rhine's work had generated. Thus, when I ran the experiment I mentioned in section 29.1, where human subjects were told to try to guess the next number a mechanical random number generator would generate, I realized that it could also be seen as an experiment in extrasensory perception (ESP). If subjects could do better than chance at this task, which was quite out of any human's control (the machine was driven by the electrical noise from a Zener diode), perhaps there was some way in which humans could have access to the subatomic world. At any rate, the disappointing, but expected, result that all subjects performed at chance accuracy dashed any secret hopes at taming parapsychology, but revealed an even more interesting behavior, another manifestation of the sequential dependencies I have been studying o¨ and on for over thirty years.

262

Chapter 29

Just before I left Duke, Rhine's foundation entertained a prominent Caribbean psychic named Lance Harribance. Many of my fellow graduate students were excited by this visit, and several went to be ``read'' by Harribance. The tales of this man's wondrous psychic abilities of course got back to the Psychology Department, and though I sco¨ed at their gullibility, I was impressed by Harribance's ability to fool these very bright students. My mentor, Greg Lockhead, was also impressed and determined to put Harribance's abilities to an objective test. He invited Harribance to our laboratory to participate in the same experiment I described above: trying to predict which numeral (1±5) would be the next one chosen by our mechanical random number generator. I had to leave before Harribance took the test, but I read all about it in Lockhead's paper (1974), unfortunately still unpublished. Supplied with food and drink, and made comfortable in our laboratory, Harribance felt con®dent that his abilities were equal to the task. He made a series of 1,000 guesses and, after each guess, he rated the con®dence he felt in the correctness of that particular guess. On each trial, he received as feedback the numeral actually chosen by the machine before making his next guess. After the test, he was even more con®dent, and felt he had done very well and had exceeded chance performance. Nevertheless (as you may have guessed), he performed very close to chance: 21.1% correct predictions, less than one standard deviation from the predicted 20%. The sequential dependencies in his guesses, however, were of exactly the same form as the ones we had described in Ward and Lockhead 1971, but the greatest in magnitude either of us had ever seen. Indeed, it was possible to predict which number Harribance would guess next with far higher accuracy than he could predict the machine's next number, using a model based on his tendency to repeat the previous machine-generated number, or choose a number at most one removed from it. When he did make such a guess, which was frequently, his con®dence was high; on the infrequent occasions when he guessed a number farther away from his last guess his con®dence was low. Harribance's feeling that he had done well came from the fact that he was within one number of the correct one on many occasions, because, as noted, small di¨erences occur most frequently in random series. By guessing a number close to the previous feedback, he ensured that even if he was incorrect, he would be close on many occasions. Unfortunately, close counts only in horseshoes and in psychic readings. In this test, rather than psychic ability, Harribance displayed only the more mundane ability to exploit a property of random sequences (cf. Goodfellow 1938; Hansel 1966).

Chapter 30 Can People Behave Chaotically?

The presence of chaotic-like human behavior is consistent with a complete determinism. . . . Yet the slightest uncertainty in knowledge makes prediction of chaotic behaviors highly unreliable. ÐAllen Neuringer and Cheryl Voss, ``Approximating Chaotic Behavior''

30.1

Yes!

Human behavior is vexingly unpredictable, and yet it is (usually) not random either (see chap. 29). Whether a ``reformed'' killer will kill again or a relationship with a person to whom we are attracted will last are questions that plague us. We need to be able to predict some aspects of human behavior and yet are notoriously bad at it. For example, it has been argued that forensic psychiatrists and psychologists are completely unable to speak with any authority about whether a particular criminal will relapse into criminal behavior if released into society, and that the testimony of such experts should not be admissible in hearings concerning such questions (Faust and Ziskin 1988). And the divorce rate is testimony that, on average, we are not good at choosing marriage partners. Where does this unpredictability come from if not from randomness? Many would answer ``free will.'' If we accept this answer, we must admit that the prospects for understanding human behavior are dim because, at a whim, a person can produce some unexpected behavior whose causal antecedents are in principle unavailable. We could possibly still discover general (nomothetic) laws having a statistical basis. For example, we could discover that, on average, 25% of rapists commit rape again within one year of release into society. However, when the question is about the behavior of a single individual rapist, does this general law say anything useful? It does only if we wish to set a criterion of 0 or 1 for the probability of a repeat o¨ense. Then we can either release (if our criterion is 1)

264

Chapter 30

or retain (if it is 0) all rapists on the basis of the ``general law.'' Otherwise, our ®nding provides no basis for saying whether any particular individual will commit the same o¨ense again. On the other hand, some researchers argue that the eventual discovery of the basis in brain activity for behavior, even covert behavior such as cognition, will allow us to predict with near certainty what an individual will do. Even better, if the individual is likely to do something we do not like, we can adjust the brain so as to prevent that behavior from emerging, a future envisaged with horror (albeit with cruder technology) by Anthony Burgess in A Clockwork Orange. Moral debates aside, the likelihood of this occurring is uncertain at present: we simply do not know enough about how the brain generates behavior and cognition. On the other hand, as the chapter epigraph indicates, chaos holds out the possibility that absolute determinism and relative unpredictability may coexist. Indeed, chaos may explain why human behavior is at the same time unpredictable and yet completely determined by discoverable laws. Although this may not help us predict future criminal behavior, where we may have to continue to rely on the dictum ``The past is the best predictor of the future,'' it might bring us philosophical solace, letting us dispense with the notion that human behavior has essentially hidden causes, or perhaps no causes at all. Chapter 26 reviewed some evidence that sequences of human behaviors might contain some nonlinear structure indicative of low-dimensional deterministic generators. Here I am going to discuss a di¨erent approach to the problem of establishing a role for chaos theory in human behavior, that of trying to teach humans to emulate a chaotic system. In a provocative article published in Psychological Science, Allen Neuringer and Cheryl Voss (Neuringer and Voss 1993) claimed they had taught humans to approximate chaotic sequences generated by the logistic di¨erence equation (eq. 24.1b): x n‡1 ˆ a
x n
…1 ÿ x n †; with a ˆ 4:0 (for a display of the strange attractor of eq. 24.1b, see ®g. 24.1B; for a brief discussion of the behavior of the logistic di¨erence equation, see sec. 24.1). Neuringer and Voss used a technique similar to the one Lockhead (1974) used to investigate the sequential behavior of a psychic (see chap. 29): they asked subjects to predict what number would occur next in a series (a one-step forecast) and then gave them feedback as to what number had actually occurred. The series in question was generated by a computer implementing equation 24.1b with seven-digit pre-

Can People Behave Chaotically?

265

Figure 30.1 One subject's data. (From Neuringer and Voss 1993. Copyright ( 1993 by Blackwell Publishers.)

cision but displaying it with three-digit precision. They also had subjects perform an analog version of this task, predicting the next location on a line that ran the width of the computer monitor. Both with lines and with digits the results were the same: subjects learned to generate forecasts that were not far from the actual numbers in the sequence based on the last number that actually occurred (the ``seed''). In ®gure 30.1, the lefthand graph shows the deviation of one subject's forecasts from the computer-generated numbers over about 4,500 trials, and the middle and right-hand graphs represent that subject's forecasts on each trial as a function of the computer-generated number on the previous trial for the ®rst and last 120 trials of this subject's data, respectively. If the subjects are behaving according to the logistic equation, then such plots should resemble ®gure 24.1B; indeed, the plot for the ®nal 120 trials does. Neuringer and Voss also subjected these data to other tests for nonlinear determinism, such as comparing the entropy (uncertainty, information) of the subjects' sequences with those of the logistic-generated sequences and with that of a random-number generator, and also ®tted their phase plots with quadratic equations. Finally, they ran an experiment in which they ceased giving the feedback at a certain point and then observed the divergence of the subjects' sequences from the computer-generated sequence, an indication of sensitive dependence (see chap. 23). On the basis of these data, they concluded that their subjects were approximating chaotic behavior, with the implications mentioned earlier. 30.2

Not Really

As I was reading Neuringer and Voss's paper (1993), I could not help thinking of several nonchaotic ways in which subjects could be generating sequences that ``approximated'' chaotic ones. People could be simply

266

Chapter 30

learning the rough shape of the logistic map, a parabola in which numbers near either 0 or 1 on the x-axis are paired with numbers near 0 on the y-axis, whereas numbers near 0.5 on the x-axis are paired with numbers near 1 on the y-axis, and so on. Enlisting the nearest family member, my daughter, I con®rmed that, after only a few trials, she could easily emulate Neuringer and Voss's subjects' behavior using only a simple heuristic that had nothing to do with chaos. Metzger (1994) had a similar thought and proposed a model in which subjects memorized several seed-feedback pairs of numbers and noticed that they fell along a parabola. Linear interpolation between these memorized pairs could then provide predictions for any other seeds. Without feedback, the subject could simply use each previous prediction as the seed for the next one. In Ward and West 1994, Robert West and I took a slightly di¨erent tack in print. We argued that from Neuringer and Voss's results it is dif®cult to tell whether subjects were iterating a chaotic equation, and thus producing a chaotic sequence and a strange attractor, or simply following a nonchaotic heuristic such as memorizing pairs of numbers and interpolating between them, when the map and the attractor are identical as they are in equation 24.1b with a ˆ 4:0. We pointed out that when a is changed from 4.0 to 3.66, although the quadratic form of the map stays the same (it is still a parabola, albeit with a slightly di¨erent pairing of numbers), the attractor of the map changes drastically, as shown in ®gure 30.2A: most of the left leg and large parts of the right leg of the map are now missing. We reasoned that if we taught subjects this map and then asked them to iterate it without feedback, they would produce the map (noisily) if they were using a heuristic such as Metzger's but the attractor if they were indeed iterating the equation they had learned, even if only approximately. As you can see from the example in ®gure 30.2B, our subjects did not produce the attractor, but only a noisy version of the map. With Metzger (1994), West and I also pointed out some subtle technical problems involving precision of computer representation of numbers for any such experiments. The crux of the matter is this: to get a good approximation of chaos from a computer, it is necessary to represent the numbers involved with at least seven-digit precision, preferably, with fourteen- or ®fteen-digit precision. Neuringer and Voss (1993), however, presented their subjects with numbers (or line lengths) having only threedigit precision, although they used seven-digit precision in the computer. With only three-digit precision, the logistic equation generates sequences that settle into limit cycles of from one to only a few points, depending on

Can People Behave Chaotically?

267

Figure 30.2 (A) Strange attractor of the logistic di¨erence equation (eq. 24.1b), a ˆ 3:66; (B) one subject's data. (From Ward and West 1994. Copyright ( 1994 by Blackwell Publishers.)

Can People Behave Chaotically?

269

the value of a, the starting number, and the rounding or truncating procedure used. Moreover, the actual seed-feedback pairings that occur in a sequence generated at seven-digit precision but rounded to three-digit precision do not properly represent even the seven-digit-precision sequence, let alone the ``real'' sequence of the logistic equation (which requires in®nite precision). On the other hand, three-digit precision is probably all that humans can comfortably deal with. In Ward and West 1994, our subjects had trouble even with that level of precision. Thus it seems that the need to use numbers of limited precision in experiments with humans calls into question the interpretation of even clear and provocative results. 30.3

Heuristics and Chaos

Despite the problems mentioned in section 30.2, it is worth investigating further how human subjects behave in the face of learning to approximate the logistic equation. Even if they are using a heuristic, it is not clear that it is the one we (Ward and West 1994) and Metzger (1994) promoted. Finding out what they are doing and how they are doing it would be of value both in understanding how people might produce unpredictable behavior, and in how they might cope with unpredictability of the form produced by chaos. To do this, West and I (Ward and West 1998) ran some additional subjects in our version of the logistic learning task. Setting the constant a in equation 24.1b to 3.66, we compared the data generated by two models to those generated by our additional subjects. One model was a computer simulation of the memory pairs model suggested by Metzger (1994); the other, a version of the logistic iteration equation model proposed by Neuringer and Voss (1993). The novel feature of our study was the analysis of both subject-generated and model-generated data using Casdagli's nonlinear forecasting technique (1991). Figure 30.3A shows a phase plot of data from a typical one of our additional subjects. It is clear that, as in Ward and West 1994, subjects failed to reproduce the strange attractor of the chaotic map they learned (®g. 30.2A). Rather they produced a phase plot that approximated the logistic map itself, in this case with added Figure 30.3 (facing page) (A) Phase plot for human subject; (B) phase plot for noisy logistic; (C ) phase plot for fuzzy memory model of subject in ®gure 30.3A. (From Ward and West 1998.)

Can People Behave Chaotically?

271

structure in the form of clumping of points in certain areas. As far as we knew, there were no such attractors in the logistic family, but we nevertheless entertained, as a model of this behavior, Equation 30.1, which is just equation 24.1b with an added term representing a sample of Gaussian (observational) noise, e n : x n‡1 ˆ a
x n
…1 ÿ x n † ‡ e n

…30:1†

When, however, we simulated this model with di¨erent amounts (standard deviations) of added noise, the resulting phase plots (e.g., ®g. 30.3B) did not contain clumping like that in the subjects' data (e.g., ®g. 30.3A). Interestingly, neither did our ®rst version of Metzger's memory pairs model (1994), in which we chose a few speci®c memory pairs for each subject based on their phase plot and then used simple linear interpolation between these pairs for other points. Although adding noise helped, in the end, we had to resort to using two types of interpolation, ``®ne'' (the usual) and ``gross,'' which consisted of choosing the midpoint between the guesses indicated by the nearest memorized pairs. We also assumed that people used fuzzy or noisy, rather than precise, representation of the memory pairs. This approach was successful at reproducing most of the features of each subject's phase plot, in particular, the clumps. Figure 30.3C is an example for the subject data shown in ®gure 30.3A. Clearly, what we called the ``fuzzy memory pairs model'' ®tted the phase plots the best. The above discussion seems to imply that subjects were not iterating the logistic equation to produce chaotic sequences of numbers, a somewhat disappointing conclusion. Nonetheless, we wanted to see whether there was any nonlinear structure in the sequences produced by both subjects and models. Perhaps the heuristic used by subjects to comply with the experimenters' instructions produced a di¨erent kind of chaotic sequence. Thus we subjected both subject-generated and model-generated sequences to the nonlinear forecasting analysis described in chapter 25. Figures 30.4A±30.4C show deterministic versus stochastic (DVS) plots of the subjects' data and the computer-generated data from each of the two models. The DVS plots of subjects' data (®g. 30.4A) show evidence of nonlinear deterministic structure: they are similar to the plots of the logistic equation with some observational noise added (®g. 30.4B). Three of the subjects' DVS plots seem well approximated by the noisy logistic Figure 30.4 (facing page) (A) Deterministic versus stochastic (DVS) plots for human subjects; (B) DVS plots for noisy logistic; (C ) DVS plots for fuzzy memory model. (From Ward and West 1998.)

272

Chapter 30

DVS plots for noise of between SD ˆ 0:05 and SD ˆ 0:10, whereas the other two required noisy logistic sequences with SD ˆ 0:20. Thus the noisy logistic model could ®t subjects' DVS plots with the choice of a single parameter, the standard deviation of the noise distribution. Second, the DVS plots of data from the fuzzy memory pairs model for the best®tting sequences (®g. 30.4C) are nearly identical to the corresponding plots for the subjects' data with one exception (subject Shir's model curve indicates less noise than the data curve). Thus the memory pairs model, which reproduces both the subjects' phase plots and their DVS plots, is a good model, but not a chaotic one. Or is it? After all, its output does show the same evidence of nonlinear structure as the subjects' plots, and could also be ®tted by the noisy logistic model (compare ®gs. 30.4B and 30.4C). We are left in a quandry, then, regarding such models, although we do know that the fuzzy memory model heuristic is e¨ective at approximating chaotic behavior. Perhaps similar heuristics could be used to produce unpredictable behavior in other situations, or even to cope with chaos itself.

Chapter 31 Relaxation Oscillators: A Foundation for Dynamical Modeling

Now that the general equation for relaxation oscillations is known, it is an easy matter to devise further electrical systems of the same type. ÐBalthasar van der Pol, ``On `Relaxation-Oscillations' '' Van der Pol's equation for a relaxation oscillator is generalized. . . . The resulting . . . model serves as a simple representative of a class of excitable-oscillatory systems including the Hodgkin-Huxley (HH) model of the squid giant axon. ÐRichard A. FitzHugh, ``Impulses and Physiological States in Theoretical Models of Nerve Membrane''

It is time, book time that is, for a change. Chapters 16 through 30 have dealt with speci®c problem areas in dynamical science: 1=f noise, stochastic resonance, and chaos theory. In the ®nal chapters, I am going to broaden the scope of the discussion to deal with more speculative topics. I hope these last ®ve chapters will be fun and their ideas worth thinking about, but I cannot guarantee that they will be even as solidly based as the admittedly incomplete work on the topics already discussed. Let us begin with a general perspective on building dynamical models. 31.1

A Brief Taxonomy of Oscillators

Time, in one way or another, is central to every chapter in this book. A dynamical system is characterized by change over time, and change over time is most compactly represented by di¨erential equations, although not all di¨erential equations are equally useful when it comes to modeling dynamical cognitive systems. One especially promising class of di¨erential equations de®nes oscillators, systems that change in some more or less regular way, and that are therefore ideal both for measuring time and for describing processes that evolve in time. A good example of an oscillator is the pendulum of a grandfather clock, which keeps reasonably good

274

Chapter 31

time because it has a roughly constant period. As we noted in chapter 4, pendulums were the ®rst precise measurement devices for time, and they have been thoroughly studied and described in all of their incarnations. Another useful example is what is called a ``simple, damped, unforced linear oscillator,'' such as the small gold ore cart with mass m (gold ore plus cart) and frictionless wheels attached to a gold mine wall via a spring with sti¨ness s, pictured in ®gure 31.1. Notice that, attached in parallel to the spring, is something called a ``dashpot damper,'' a container ®lled with oil and a piston that tends to act against the force exerted by the spring on the cart. Imagine we have attached a mule to the cart and we make the mule pull the cart away from the wall to which the spring is attached. When the mule begins to complain that the load is getting heavy, we release the cart. It rushes back toward the mine wall until the spring is compressed somewhat, at which point it is pushed away again, and so on, oscillating back and forth until it ®nally slows to rest, its oscillations ``damped out'' by the dashpot damper, which provides a force opposing both inward and outward movements. Although hardly a good way to go about getting gold ore out of a gold mine, this system does have some useful properties, which physicists and engineers have exploited. A particular variant of it, called a ``relaxation oscillator,'' has the potential to form the foundation for modeling a wide class of dynamical cognitive systems, including the brain itself. The typical di¨erential equation that describes the motion of an oscillator such as that shown in ®gure 31.1 is m

d 2x dx ‡ r ‡ sx ˆ 0; 2 dt dt

…31:1†

where x is position, t is time, m and s are mass and spring sti¨ness, as mentioned, and r represents the amount of damping provided by the dashpot. Here d 2 x=dt 2 is simply a compact way of writing d…dx=dt†=dt, the rate of change of the rate of change of position, or acceleration (formally called the ``second derivative of position with respect to time''), which we ®rst encountered in section 4.1. Typically, equation 31.1 is divided by the mass, m, to put it in standard form: d 2x dx ‡ b ‡ cx ˆ 0; dt 2 dt

…31:2†

where b ˆ r=m and c ˆ s=m. This equation describes a spiral path in phase space (left side of ®gure 31.1) as the amplitude of the sinusoidal,

Figure 31.1 Simple damped, unforced, linear oscillator. (From Thompson and Stewart 1986. ( 1986 by John Wiley & Sons, Ltd. Reprinted by permission.)

Relaxation Oscillators: A Foundation for Dynamical Modeling 275

276

Chapter 31

back-and-forth motion of the ore cart gradually decreases until the cart ®nally comes to rest (right side of ®gure 31.1). Oscillators in general can be categorized with respect to the three attributes just mentioned: linearity, damping, and forcing. Di¨erential equations, for their part, can be divided into two large groups: linear and nonlinear. Linear di¨erential equations have all terms of either degree one or degree zero (degree refers to the exponents of the derivatives). Equations 31.1 and 31.2 are linear because the derivatives in their terms all have degree one (including the ®rst term, which, as we noted represents the second derivative, not the square of the ®rst derivative). Squaring dx/dt, among other possibilities, would make them nonlinear. Linear differential equations exhibit some very special properties (symmetries) that often allow them to be solved exactly, whereas nonlinear equations lack such symmetries and are very di½cult to work with. The real world is always nonlinear but often can be approximated usefully by linear equations. Damping refers to a force opposing the ``natural'' oscillation of the system. Without the dashpot damper, the system portrayed in ®gure 31.1 would continue its sinusoidal oscillations forever at the same amplitude. The damping in equation 31.2 is linear because the damping term, b…dx=dt†, is proportional to the velocity of the cart, (dx=dt). If it were proportional to the square of the velocity the damping would be nonlinear. Oscillators in the real world are always damped; that is, they are not perfectly e½cientÐthey lose or dissipate energy. Sometimes the damping can be negligible over the given time period, but usually it is not. Finally, forcing refers to energy inputs from outside the given system. If our mule pulled at and pushed at the cart repeatedly even as it oscillated under the in¯uence of the spring (which stores and then releases the energy of the initial displacement), the oscillation would be forced. In the real world, oscillations are always forced because no system can be completely isolated from the rest of the universe. The forcing can, however, sometimes be so small that it can be neglected. The oscillators with the most interesting behavior are both damped and forced. The interaction of damping and forcing with the basic structure of a system can generate complex dynamicsÐoften chaoticÐand a complicated phase space picture, showing qualitatively di¨erent behaviors for di¨erent parameter regimes. Thompson and Stewart (1986) describe in some detail the various types of oscillators that can be generated from various combinations of linearity, damping and forcing. Although a number of these oscillators could be applied in dynamical cognitive science, either to model the temporal unfolding of cognitive phenomena or to measure systemic time,

Relaxation Oscillators: A Foundation for Dynamical Modeling

277

section 31.2 concentrates on one particular type of nonlinear oscillator that I believe could become a foundation stone of dynamical cognitive science. 31.2

The van der Pol Relaxation Oscillator

Let us ®rst consider the van der Pol equation, which describes an unforced, damped, nonlinear oscillator: d 2v dv ‡ a…v 2 ÿ 1† ‡ o 2 n ˆ 0; dt 2 dt

…31:3†

where the variable v could stand for the voltage across a resistance in an electrical circuit. In his initial studies of the equation, van der Pol (1926, 992) suggested several such circuits, of which the simplest was a neon lamp in parallel with a capacitor, this pair in series with a resistance and a battery, and ``perhaps also heart-beats.'' Notice that the damping term (the second term) is what makes the equation nonlinear: because the function a…v 2 ÿ 1† multiplies the ®rst derivative (dv=dt), this term is of third degree. The relationship between a and o is crucial to the behavior of the oscillator. When a f o, the oscillator behaves like a linear oscillator with a very small amount of damping, generating a very slowly decaying sine wave, as portrayed in ®gure 31.2 (top panel). When a A o,

Figure 31.2 Behaviors of the van der Pol equation (eq. 31.3) for o ˆ 1. (From Thompson and Stewart 1986. ( 1986 by John Wiley & Sons, Ltd. Reprinted by permission.)

278

Chapter 31

the oscillatory behavior becomes less sinusoidal but is still fairly regular (middle panel). When, however, a g o, the oscillation approaches a square wave (bottom panel). Notice in the bottom tracing of ®gure 31.2 that each sudden transition (nearly vertical trace) is preceded by a considerable time period during which the voltage changes only slowly. Van der Pol (1926) called this phenomenon a ``relaxation oscillation'' because each slow change followed by a quick jump resembles the buildup and release of charge in a capacitor with capacitance C and ``relaxation time'' t ˆ RC (where R is the resistence of the circuit), the time it takes for the capacitor to charge. In this system relaxation, or the time taken for the ``memory'' of the previous state to decay, is the important feature, not the restoring force (o 2 v). A relaxation oscillator is highly prone to phase locking with an external driving frequency (forcing), even one that does not correspond to the natural frequency of the unforced oscillations, all the while maintaining a relatively constant oscillation amplitude. These properties make such oscillators ideal systems for controlling a system that should produce a response of ®xed amplitude but variable frequency, for example, the heart or a neuron. Recognizing this, van der Pol and van der Mark (1928) made an analog model of the electrical activity of the heart consisting of three coupled relaxation oscillators: one to generate sine wave forcing of the auricles, a second to represent the auricles, and a third to represent the ventricles (contraction of the auricles drives that of the ventricles). This model explained (qualitatively) both the normal heartbeat, including the ``all-or-nothing law'' and various irregularities, and some disorders of the heartbeat such as heart block, in which decoupling the auricle from the ventricle leads to heart failure. Relaxation oscillators can even give rise to chaos. In another famous paper, van der Pol and van der Mark (1927) reported that the neon tube oscillator they were to use later as a model of the heart had a remarkable property. If a forcing oscillator was added to the circuit in series with the neon tube, the nonlinearity of the relaxation oscillator made the whole circuit produce oscillations at frequencies that were equal to the forcing frequency divided by various whole numbers (2, 3, 4, 5, 6), the whole number increasing as the parallel capacitance was increased. This phenomenon, illustrated in ®gure 31.3, turned out to be very useful for electronic engineering. But the two researchers also noted an anomaly in this circuit: at certain values of the capacitance, ``often an irregular noise is heard in the telephone receivers before the frequency jumps to the next lower value'' (p. 364), a phenomenon represented by the hatched areas in ®gure 31.3, mostly at higher capacitances but at lower ones as well. This

Relaxation Oscillators: A Foundation for Dynamical Modeling

279

Figure 31.3 Chaos in a relaxation oscillator. (From van der Pol and van der Mark 1927. ( 1927, Macmillan Magazines, Limited. Reprinted with permission from Nature.)

noise was chaos (Kellert 1993 also noted this). Note that the equation describing this circuit is equation 31.3 with a simple sinusoidal forcing term added (E0 sin ot), thus completely deterministic. Moreover, although there was certainly random circuit noise present, it was so small that it does not show in the graph at the middle values of capacitance. The ``irregular noise'' only showed up just as the system was switching to a di¨erent oscillatory mode (attractor). Finally, note also the hysteresis diagrammed in ®gure 31.3 at a capacitance of about 20 mF. Hysteresis is also an important signature of nonlinearity, in this case indicating the importance of knowing the path by which a system has reached a certain state, yet it was not even mentioned by van der Pol and van der Mark in their 1927 paper. Relaxation times in physical systems correspond to ``memory'' in cognitive systems. That is, a ``relaxation'' is a time period during which the in¯uence of previous events on current behavior gradually declines. In a time series, this is measured by correlations between current behavior and behavior at various lags. Human behavior is characterized by time series that generate rhythms, orders, sequences, and so forth, even chaos. The relaxation oscillation is a formal construct, represented by the behavior of

280

Chapter 31

equations such as equation 31.3, that could be used to model the ubiquitous temporal correlations found in human cognitive and other behavior. Because such relaxations can happen at several scales in human behavior, coupled systems of relaxation equations would be required, making the task of analysis very di½cult (even the analysis of equation 31.3 is very di½cult). Several such systems have been described in chapters 16, 18 and 19 in the context of 1=f noise, as well as in the chapters on chaos theory. Section 31.3, describes how the van der Pol equation can be applied to the oscillations of neurons in the brain. 31.3

Noisy Oscillators in the Brain

Neurons display relaxation time behavior par excellence: they ®re an action potential (spike) when the voltage across the cell membrane in the trigger zone of the soma (around the base of the axon that leaves the soma) exceeds a threshold, and then ``gradually'' build back up again, usually under the in¯uence of input from other neurons, until the threshold is crossed again and then they ®re another spike, and so forth. Hodgkin and Huxley (1952) developed an in¯uential mathematical model of this process based on studies of action potentials in the squid giant axon. The model consists of four ®rst-order di¨erential equations that describe the ¯ow of ions through the cell membrane arising from several physiological mechanisms (e.g., the sodium pump, passive di¨usion). The equations are not easy to solve or analyze, and they have been the subject of much study, even though they do not produce all of the phenomena shown by vertebrate neurons (for example, they do not show adaptation to a constant stimulus; Yamada, Koch, and Adams 1998). Richard FitzHugh (1960) was one of those who worked on the analysis of the Hodgkin-Huxley equations. He ®rst tried describing their behavior using phase space methods, such as those discussed earlier in this book, combined with reducing the system of equations to a more manageable size by holding one or more of the variables of state constant, that is, by assuming that their derivatives were equal to zero. Finding that approach inadequate to describe how trains of impulses occur, he noticed that neural action potentials are actually relaxation oscillations similar to those described by van der Pol and van der Mark (1928) for the heartbeat. In an in¯uential paper, FitzHugh (1961) described a model of a neuron based on a special case of equation 31.3 (o ˆ 1) plus some additional terms, including a forcing input, and showed how it mimicked the behavior of the HodgkinHuxley equations. The resulting equations were studied further by Na-

Relaxation Oscillators: A Foundation for Dynamical Modeling

281

gumo, Arimoto, and Yoshizawa (1962); they have come to be called the ``FitzHugh-Nagumo model'' (although FitzHugh modestly called it the ``Bonhoe¨er-van der Pol model''). The FitzHugh-Nagumo equations, in one popular form, look like this: dn ˆ a…w ‡ n ÿ n 3 =3 ‡ z† dt

…31:4†

dw ˆ ÿ…n ÿ a ‡ bw†=a; dt

…31:5†

where v is a voltage across the neuronal membrane that represents a fast, excitation process, w is a voltage that represents a slower, recovery process, z is the input stimulus intensity (the forcing term representing input from other neurons), a is the damping constant (greater than o ˆ 1, so that the system shows relaxation oscillations), and a and b are constants that a¨ect the recovery rate. FitzHugh (1961), in showing that this system mimicked the Hodgkin-Huxley equations, also showed that the neural action potential is a form of relaxation oscillation. Since 1961, the FitzHugh-Nagumo model of the neuron has been used in many biophysics applications, especially in the study of stochastic resonance in excitable systems (see chap. 22), where stochastic (noise) forcing is added to the ``fast'' equation 31.4 and deterministic forcing is added to the ``slow'' equation 31.5 (Longtin 1993). Such additions are consistent with the usual case that the noise is at a much faster timescale (comparable to spike duration) than is the signal (comparable to recovery duration). Longtin's equations (1993) look like this: e

dn ˆ n…n ÿ a†…1 ÿ n† ÿ w ‡ x…t† dt

dw ˆ n ÿ cw ÿ b ÿ r sin…bt† dt

…31:6† …31:7†

where x…t† is the noise forcing (an Ornstein-Uhlenbeck process; see chaps. 14±15) and r sin…bt† is the deterministic forcing (replacing the z in eq. 31.4). (There are also a few other changes in equations 31.6 and 31.7 from equations 31.4 and 31.5, but they do not change the essence of the process.) Longtin (1993) showed that these equations behave much like neurons (see ®g. 31.4) and also demonstrate stochastic resonance (see chap. 22). In further papers, he has shown how the equations' behavior can be synchronized to periodic forcing (Longtin 1995a) and that it demonstrates stochastic phase locking (Longtin 1995b).

282

Chapter 31

Figure 31.4 Sequence of spikes from Longtin's Fitzhugh-Nagumo model of a neuron. (From Longtin 1993.)

Many more examples of the relevance of relaxation oscillators to dynamical cognitive science can be adduced, including the double potential well Haken, Kelso, and Bunz (HKB, 1985) model discussed in chapter 3. This model can be generalized to as many wells as necessary (even in®nite), to model ever more complex brain systems. I am suggesting that systems of coupled relaxation oscillators (neurons) at many scales are su½cient to produce rhythm, timing, serial order, and many of the other phenomena of the unfolding of human behavior in time, including the recurrence of thoughts analyzed by Crovitz (1970) and the stream of consciousness described by James (1890). Indeed, modeling such phenomena with relaxation oscillators allows us to understand the sense in which such dynamical systems create their own timeÐthe system itself is oscillating, and oscillations are the only means we have of measuring time. In a deep sense, oscillations are time. Models based on relaxation oscillators would explain why the behavior emerges ``when'' it does, and why there are correlations between behaviors at many scales, including possibly chaotic behavior. Stochastically driven and deterministically driven relaxation oscillators comprise stochastic processes such as those described in chapters 12 and 13, and some of the modern tools of mathe-

Relaxation Oscillators: A Foundation for Dynamical Modeling

283

matics and physics can be brought to bear on their analysis. Such models should be particularly well suited to describe the synchronization of neurons and the temporal coding of information thought to underlie much neural activity (e.g., Singer 1993, 1994) and possibly also to form the basis of human consciousness (e.g., Tononi and Edelman 1998; see chap. 35). Already the in¯uence of relaxation oscillators in neural modeling is pervasive (see, for example, Koch and Segev 1998). The promise is that modeling human cognitive phenomena with relaxation oscillators will allow current and future neural models to be coupled to behavioral models, bringing dynamical modeling to a new level of complexity and usefulness.

This page intentionally left blank

Chapter 32 Evolution and Ecology of Cognition

The half-art, half-instinct of language still bears the stamp of its gradual evolution. ÐCharles Darwin, The Descent of Man Animal cognition is a biological feature that has been molded by natural selection. Hence, to understand cognition, one must study its ecological consequences and evolution. ÐReuven Dukas, Cognitive Ecology

32.1

Evolution of Cognition

Thus far, we have considered dynamical cognitive systems at many timescales, from milliseconds to seconds, minutes, hours, even days. Although ``historical'' timescales, such as years, decades, centuries, and millenia, are highly relevant to understanding and predicting human behavior and there is every reason to believe dynamical models will illuminate human behavior on such timescales, that is the work of disciplines other than dynamical cognitive science. The very longest timescales, however, those on which the processes of evolution work, are directly relevant to the topic of this book. Evolutionary theory can often provide the answers to ultimate ``why'' questions, both about morphology (what an animal is like physically) and about behavior. For example, evolutionary theory sheds light on the mysteries of the male peacock's tail and doves' courtship rituals, among dozens of other profound questions (see Lea 1984 for an excellent introduction to this function of evolutionary theory in understanding behavior). It has the promise to do the same for questions about cognition, such as why nonhuman animals exhibit selective attention and short-term memory but not explicit, episodic memory, or why bacteria do not have (or need) short-term memory. Evolutionary theory also provides a framework within which to appraise current systems.

286

Chapter 32

There is even a new ®eld of psychology, called ``evolutionary psychology,'' that takes this approach to the study of human behavior (see Buss 1999 for a survey). Of course biology has been studying animal behavior in this way for many years, and Nobel prizes have been awarded to Karl von Frisch, Niko Tinbergen, and Konrad Lorenz for the development of the methods and early results of comparative ethology. However, both evolutionary psychology, which emphasizes the study of human mating strategies, and comparative ethology, which emphasizes the study of aggression and reproduction in nonhuman animals, have neglected what Dukas (1998a) calls ``cognitive ecology,'' the study of the evolution of cognitive capacities and their in¯uence on animal-environment interactions, even though such a study was clearly mandated by Darwin (see ®rst epigraph above). This chapter discusses a few of the possibilities for dynamical cognitive science engaging with evolutionary timescales. Darwin's great theory of evolution by natural selection was based on two simple observations and a conclusion that followed ``naturally'' from them. First, within each species there is quite dramatic variation of structure, even though all members of the species share certain characteristics. For example, although each human has a head with two eyes, a nose, a mouth, hair, and so on, human heads come in all sorts of shapes and colors, and so do their parts. A similar statement can be made for every characteristic of every animal, although the amount of variation can be more or less than that of human heads. Moreover, Darwin observed that this variation tends to be inherited. Short parents tend to have short children, muscular parents tend to have muscular children, and so forth. Second, every species produces many more o¨spring than could possibly survive. Darwin calculated for many species how long it would take for an unsustainable number of its members to be born, at species-typical reproductive rates and assuming survival of all o¨spring. It was a surprisingly short time, even for the slowest reproducers. For example, a female cod lays millions of eggs each time she reproduces, which can happen several times a year. If each egg grew into an adult, and survived and reproduced in turn, the oceans would literally be over¯owing with cod after only a few months. From these two observations, he concluded that if living creatures vary and not all survive, the survivors must possess the inheritable characteristics that are best adapted to the local environment in which the creature must live. This principle of natural selection was called the ``survival of the ®ttest'' because it was originally thought to amount to a struggle between individuals to see who would, literally, live on from day to day. But gradually, it has come to be interpreted in a

Evolution and Ecology of Cognition

287

more useful way, as a struggle for successful reproduction. Of course an organism that does not survive to reproductive age cannot reproduce, and its genes are lost to the species. Thus the struggle to live long enough to reproduce is part of natural selection. But there are also more subtle factors at work, such as which stag can attract and hold the largest harem of does against competing stags, or which male peacock can attract the locally most fecund female with his magni®cent plumage. Today, ®tness is de®ned simply as the ability to produce viable o¨spring, and it is the universal yardstick of evolution; the genes that persist in a species are the ones that maximize the ®tness of their possessors. It is simply a numbers gameÐDarwinian roulette. Because evolution by natural selection (with modern revisions and with additions such as population genetics; see especially Luria, Gould, and Singer 1981) is one of the two theoretical cornerstones of biology (the other is the ``central dogma'' of molecular biology on the pivotal role of DNA), I cannot possibly survey it here. I can, however, point out a few aspects relevant to dynamical cognitive science. One of these has to do with the origin of traits. Although Darwin concentrated on explaining the origin of species (albeit with only modest success), the general approach of evolutionary theory can also usefully be applied on a smaller scale to the understanding of the origin of traits. For example, senescence has been usefully analyzed, and its role in the overall functioning of a species in its niche elucidated, from this point of view. Similarly, many behavior traits have been successfully analyzed in this way, especially mating rituals among nonhuman animals. A similar analysis should reveal how information processing abilities evolved among animals, from the simplest one-celled animals such as bacteria, whose cell membrane responds directly to chemical gradients, to the most complex human accomplishments such as performing a piece of music or creating a scienti®c theory. Methods of comparative ethology could be adapted to deduce evolutionary trends in, for example, memory use, or communication techniques. Although some have begun to study cognition in this way, there is still much to be learned about how and why the ability to write a book such as this came to be among those abilities possessed by humans. Of course the resulting picture is not likely to be simple. One complication is that collections of traits could have coevolved. For example, it has been suggested that, because speech motor patterns seem to be the substrate for speech perception (e.g., Liberman and Mattingly 1985), the mechanisms for the perception of and for the production of human speech must have coevolved. Another possibility is a mutation of a ``regulator

288

Chapter 32

gene.'' Ordinary genes code for only a single trait, and sometimes only for a part of a trait. Regulator genes, on the other hand, do just what the name suggestsÐthey regulate the expression of many genes, in turn in¯uencing many traits. One or more such mutations could have been the change that led to the emergence of humans (see Stanley 1981 for a development of this theory). Similarly, mutation in a regulator gene that governs expression of the genes that code for brain structure could have simultaneously a¨ected several information-processing mechanisms together, producing an animal that had both enhanced visual pattern recognition abilities and the ability to focus them on particular patterns in the visual ®eld. Another complication is that the modern view of evolution has replaced Darwinian gradualism (many small changes over many years) with a more interesting dynamical process, called ``punctuated equilibrium,'' in which things are relatively stable for millions of years and then something dramatic happens, either an explosion of new species or an extinction of huge numbers of existing species (see again Stanley 1981). This represents a kind of oscillation of species production within a time series lasting billions of years. Looking at evolution this way focuses on macroevolutionary trends, for example, the way that the appearance of new species changes the distribution of traits in the world. Evolution is essentially historical, and cannot create traits out of whole cloth: it works with what is available. Macroevolution, on the other hand, works with the species available, and can be responsible for dramatic changes that have nothing to do with individual ®tness but only with the corresponding concept de®ned at the species level, which involves both how long a given species persists and how fecund it is with respect to giving rise to o¨spring species. An example of how this could work is displayed in ®gure 32.1. Here a hypothetical evolutionary change in shape portrayed in the top graph takes place simply because species tend to last longer (middle graph), and to give rise to more new species (bottom graph), the farther they are to the right on the ``shape'' axis. Thus either micro- or macroevolution or both could be involved in producing changes in information-processing abilities of animals. Moreover, the long-term oscillations of macroevolution can also be treated with the same tools that we have developed for dealing with shorter-term changes. Figure 32.2 shows how the logistic equation (eq. 8.6) can be ®tted both to the growth of a bacterial colony over a period of several hours and to the growth in the numbers of marine orders (collections of species) over hundreds of millions of years.

Evolution and Ecology of Cognition

289

Figure 32.1 Macroevolutionary trend. (From Stanley 1981. ( 1981 by Steven M. Stanley. Reprinted by permission of Basic Books, a member of Perseus Books, L.L.C.)

Although the Darwinian approach to evolution has been modi®ed in many ways, one of its sacred cows is the idea that evolution is ``blind,'' that it does not proceed by design but rather by chance, that it does not always lead to ``progress.'' This idea is eloquently defended by Dawkins (1986) in his book The Blind Watchmaker. Moreover, as explained by Dawkins (1976) in his earlier book, The Sel®sh Gene, much if not most evolution is at the level of the individual organism, where the genes have only self-interest. Evolution is not supposed to be for the good of the organism's own species, much less than for the good of other species. On

Figure 32.2 Time series at two time scales conform to the logistic equation. (From Luria, Gould, and Singer 1981. ( 1981 by The Benjamin/Cummings Publishing Company, Inc. Reprinted by permission of Addison Wesley Longman, Inc.)

Evolution and Ecology of Cognition

291

the other hand, this and other evolutionary sacred cows have also been challenged, nowhere more forcefully than by Stuart Kau¨man (1993) in his profound but di½cult work The Origins of Order. Kau¨man argued that complex systems poised at the ``edge of chaos'' are the most able to adapt and change through the processes of mutation and natural selection. The edge of chaos is a place in the phase space of a complex dynamical system where the orbit is highly complex but not quite chaotic. Typically, only a small change in one or more parameters is necessary to push the system into chaos. For example, the logistic di¨erence equation (eq. 24.1a) iterating with a value of the parameter a ˆ 3:57 would be at the edge of chaos; if a b 3:58, the attractor can be chaotic. Kau¨man also argued that coevolving complex systems could be pushed by selection to form ecosystems where all species mutually attain the edge of chaos, and thus symbiotically maintain the maximum possible ®tness for all. If Kau¨man's arguments are correct, then our own information processing mechanisms could have evolved mutually with those of the other animals around us, to the mutual bene®t of all. If so, we cannot understand our own mechanisms without also understanding those of our coevolvers. 32.2

Ecology of Cognition

How an animal obtains and processes information about its environment clearly constrains that animal's interactions with that environment, its ``cognitive ecology,'' which in turn determine an animal's ®tness: the survival of its genes into future generations. And that ®tness, in turn, determines which of those information-processing capabilities will be present in those future generations. The same selection pressures that brought about the male peacock's gorgeous (but heavy) tail plumage are responsible for such things as selective attention, implicit and explicit memory, classical and operant conditioning, language, and so forth. (For useful discussions of the evolutionary perspective and the study of particular cognitive constraints on animal behavior within a particular ecological niche, see Dukas (1998c.) Dukas's introduction and chapter (1998a,b) contain much of the material most relevant to dynamical cognitive science. In his chapter (Dukas 1998b) on the e¨ects of information-processing constraints on animal behavior, Dukas sees selective attention as an adaptation to the requirement of some niches for highly demanding computations to accomplish predator and prey recognition. The focus of attention provides a restricted range over which such computations can be performed

292

Chapter 32

e½ciently with a limited neural apparatus, albeit with associated costs to detection of predators or prey in parts of the environment outside the attentional focus. One cost of attentive focus is that it requires e¨ort, e¨ort that, when directed at focusing continuously on a di½cult task, results in an increasing decrement in performance the longer the task lasts. Since its discovery among radar monitors during World War II, this vigilance decrement has been studied extensively in humans. Attentive focus is required by animals during foraging, a ubiquitous and vitally important activity of all animals, or by humans during the typical workday (which has replaced foraging for many of us). Satisfactory performance during foraging or working in an o½ce, factory, or store demands occasional rest periods for recovery from vigilance decrement (among other decrements). This observation suggests that an animal should alternate (oscillate) between foraging and rest periods during its foraging activity rather than simply foraging until it is satis®ed. Dukas and Clark (1995) proposed a dynamical model of foraging that has just this character. First, they assumed that a foraging animal's vigilance level declines linearly during active foraging according to equation 32.1: dn ˆ ÿan; dt

…32:1†

where a is the rate of vigilance decrement, which is associated both with task di½culty (larger, the more di½cult the task) and with the individual animal. Dukas and Clark also assumed that the forager's vigilance level recovers during a rest period according to dn ˆ b…1 ÿ n†; dt

…32:2†

where b is the rate of recovery, associated both with the quality of rest the animal is able to attain and with the individual animal. These two linear di¨erential equations can be combined into one equation by assuming that the foraging animal devotes some fraction, 0 < y < 1, of its time to foraging, and the remaining amount of time, 1 ÿ y, to resting. In this model, y re¯ects the proportion of its time the animal puts into foraging, thus its foraging e¨ort. Forming a linear combination of equations 32.1 and 32.2 with weights y and 1 ÿ y, respectively, for foraging and resting, results in dn ˆ ÿyan ‡ …1 ÿ y†b…1 ÿ n†: dt

…32:3†

Evolution and Ecology of Cognition

293

Figure 32.3 Optimal length of a foraging bout. (From Dukas 1998, p. 109. ( 1998 by The University of Chicago.)

One (not so natural) way to use this equation is to assume that a forager alternates rapidly between periods of foraging and periods of rest and thus maintains an equilibrium level of vigilance, n…y†. This equilibrium level of vigilance can be found by setting dv=dt equal to zero and solving for v. The result is n…y† ˆ

…1 ÿ y†b : ya ‡ …1 ÿ y†b

…32:4†

Several things can be discovered by contemplating (or even better, graphing) this equation for various values of a; b; y, and that for average food intake, expressed by f …y† ˆ lyn…y†:

…32:5†

One thing we discover is that equilibrium vigilance level is lower, the greater the ratio a=b. Thus di½cult tasks such as foraging for cryptic (well-camou¯aged) prey, which produce greater rates of vigilance decrement (higher a), also produce lower equilibrium levels of vigilance. Moreover, although increases in foraging e¨ort (increases in y) result at ®rst in increases in food intake when y is small (eq. 32.5), when y is large, such increases actually decrease food intake because of the simultaneous decrease in the equilibrium vigilance level. Thus, for any ratio a=b, there is an optimum proportion of time to spend foraging, and that proportion increases the easier the foraging task is. This somewhat counterintuitive conclusion is typical of the results one can obtain from such dynamical models even when treated statically (at equilibrium). A more dynamical way to view this model, in which it was assumed that a forager would initiate foraging activity when it was well rested, was

294

Chapter 32

also discussed by Dukas and Clark (1995). The animal's vigilance level would start at v1 and then decrease with time (exponentially, because this is a solution of the simple linear di¨erential eq. 32.1) to some level v0 , at which time it would be fatigued and would begin to rest. They also assumed that there would be a time delay, t, after foraging before the animal could actually begin to rest (a cost of the switch in activitiesÐ perhaps the time it takes the animal to ®nd a predator-free refuge). Then, once resting, vigilance would build back up again, logarithmically, to the rested level, v1 . Figure 32.3 shows one foraging cycle. Clearly, if the rate of decrement of vigilance, a, is large, then the optimal foraging bout will be short, and if the cost of switching activities, t, is large, then foraging bouts also should be short. In this linear model, an animal's days consist of a sequence of alternating periods of foraging and resting, foraging and resting, with perhaps an occasional bout of running from a predator, until death. Of course in this model the foraging and rest periods are of ®xed length, and thus would not model any animal's actual life. If, however, we introduce some nonlinearities, or some noise, or both, then the alternations become more irregular, and thus more lifelike. Notice that if we extend the single cycle of ®gure 32.3 to an inde®nite sequence of such cycles, including the irregularities, they very much resemble the cycles of a relaxation oscillator. In fact, a very similar picture could be produced using a van der Pol relaxation oscillator. In such a nonlinear, dynamical model, however, the transitions between foraging and resting would be more sudden, and the curves of decreasing vigilance and increasing capacity would have di¨erent shapes (see ®g. 32.2). Again, introducing noise into this picture would create the irregularity observed in nature and also perhaps create the rare departures from a regular pattern. In such a model, the parameters would be the amount of damping (equivalent to the rate of vigilance decrement) and the restoring force (equivalent to the rate of regaining attentional capacity). A similar model could be proposed for humans to describe both performance in vigilance situations and the rhythm of work and rest during a typical day. Whether the original linear model proposed by Dukas and Clark (1995) would be su½cient to account for the many di¨erent aspects of vigilance behavior, especially when applied to human activities, or the nonlinear model just sketched out would be required, is left for future researchers to ascertain. What is clear, however, is that looking at the dynamics of cognition in the context of the demands of survival in a biological niche during evolutionary times promises many insights not attainable from a statical view of cognitive capacities.

Chapter 33 Dynamical Cognitive Neuroscience

In studying the relationship of mind and brain . . . neuroimaging techniques are of paramount importance, for only these techniques can provide pictures of the active, living human brain. ÐMichael I. Posner and Marcus E. Raichle, Images of Mind Neocortical dynamic behavior at multiple spatial scales, ranging from molecules to neurons to overlapping local and regional cell groups of di¨erent sizes to global ®elds of synaptic action density. Interaction across these hierarchical levels (or spatial scales) may be essential to the dynamics (and by implication to behavior and cognition). . . . ÐPaul L. Nunez, ``Toward a Quantitative Description of Large-Scale Neocortical Dynamic Function and EEG''

To study the dynamics of the brain, we must measure brain activity. For many years such measurement was impossible. In the mid-1800s, eminent scientists like Gustav Teodor Fechner, the founder of psychophysics, believed that we would never be able to study ``inner psychophysics,'' the relationship between brain activity and the mind. This belief stimulated him to develop ``outer psychophysics,'' the study of the relationship between the world of physical stimuli and the mind, or at least conscious sensations. Nonetheless, beginning in the late 1800s, scienti®c and technical developments combined to produce a plethora of techniques now used to study brain activity (see Posner and Raichle 1994 for a nontechnical introduction and Kandel, Schwartz, and Jessell 2000 for a technical survey of the state of our current knowledge of the brain). Many of these, such as inserting electrodes directly into the brain, are quite invasive and, as such, can only be performed on nonhuman animals or used on human patients when the information they provide will help with diagnosis or therapy. In recent years, however, several less invasive techniques have emerged and are beginning to provide very useful information about

296

Chapter 33

brain function in intact, normal humans. Some of these techniques have been mentioned already: magnetoencephalography and electroencephalography are particularly suited to providing time series of brain activity with millisecond accuracy for time periods of up to many minutes. Other techniques, depending on the fact that more neural activity requires more blood ¯ow to replenish oxygen and energy used in generating action potentials and other neural activities, can provide more precise spatial localization of neural activity associated with particular cognitive functions. Together these techniques promise to illuminate the relationship between brain and mind to an extent undreamed of when Fechner lived. This chapter brie¯y describes some of the techniques and their potential role in dynamical cognitive science, conveying the ¯avor of current dynamical brain models, which involve various nonlinear oscillators such as those described in chapter 31. 33.1

Brain Imaging

We have known since Galvani made a frog's leg twitch that electrical currents run in our bodies. Nowhere is this more true than in our brains, where the basic mode of communication between neurons involves the exchange of chemicals, the neurotransmitters, that a¨ect the permeability of the neurons' cell membranes to ions, which are particles that carry electrical charge. The ¯ow of electrical charge creates electrical currents, and when enough neurons, aligned in parallel, are active, the resulting electrical currents can become quite largeÐlarge enough to reach the surface of the scalp and be detected there by special electrodes. Recording electrical ¯uctuations at the scalp, or in special cases from the dura of the brain (the sheath of the brain beneath the skull), called ``electroencephalography,'' was invented by Hans Berger in the early 1930s. The electroencephalogram (EEG) can provide long (up to several hours), precise (millisecond accuracy) time series of the global (arising from many neurons) electrical activity of the brain. A standard EEG simply records the activity from several sites on the scalp while an animal or a person does various activities. A set of such tracings from a rat brain is shown in ®gure 33.1. Notice how the oscillatory trace varies with the rat's activities and experiences. In particular, notice that there are several scales of oscillations in these tracings, and that the more global ones are more closely associated with moment-to-moment changes in behavior, such as moving while sleeping, jumping, turning the head, chattering the teeth, whereas the more local ones seem to di¨er more when the animal is in

Dynamical Cognitive Neuroscience

297

Figure 33.1 Electroencephalogram (EEG) time series of rat brain electrical activity. (From Whishaw and Vanderwolf 1973.)

di¨erent general behavioral states, such as sitting, sleeping, or swimming. Various frequencies of EEG waves are associated with particular states or activities. For example, the ``alpha'' frequency of about 10 Hz indicates a state of relaxation, whereas the ``gamma'' frequency of about 40 Hz is associated with strenuous mental activity. EEGs can be used to de®ne various behavioral states, such as the stages of sleep or relaxation versus stimulation. Because the time series of electrical activity is a¨ected noticeably by the abnormal brain tissue, comparisons of the EEG tracings of normal people with those suspected of a disease (such as a brain tumor) can often provide valuable diagnostic information. Chapter 27 describes how the standard EEG can be used as an index of chaos in the brain. Another way to use the EEG is to record the less regular ¯uctuations in brain electrical activity associated with the perception of external stimulus events (see Picton, Lins, and Scherg 1995). Such event-related potentials

298

Chapter 33

(ERPs) provide time series over the range of a few milliseconds to perhaps a second or two. Because ERPs are very small in relation to the overall activity of the brain, usually many of them are averaged to provide a clean record. When this averaging is done, speci®c features in the time series can be linked to speci®c cognitive operations, such as processing a visual stimulus in the occipital cortex, where di¨erences, say as a function of whether the visual stimulus was attended to or not, can be detected as di¨erences between ERPs under the two conditions. These di¨erences provide converging evidence to model cognitive processes (see Hillyard and Picton 1987 or NaÈaÈtaÈnen 1992 for more). Because their vulnerability to noise makes study of individual, unaveraged ERPs di½cult, it would seem to be impossible to investigate ERPs at longer timescales, but there are ways around this problem. Chapter 22 describes the search for 1= f noise in time series of individual ERPs; chapter 27 describes an attempt to ®nd chaos in ERPs; and there are now techniques available to ®lter the noise from individual ERPs so that time series of ``cleaned-up'' ERPs can be generated (e.g., JasÂkowski and Verleger 1999, 2000). This should allow ERP and behavioral time series to be coordinated. Electrical currents generate magnetic ®elds; those generated by the electrical currents ¯owing in the brain can be detected and recorded with a technique more recent than electroencephalographyÐmagnetoencephalography. Using an expensive and cranky device called a ``supercooled quantum interference device'' (SQUID), the magnetoencephalogram (MEG) provides time series of magnetic ®eld measurements across the brain that are possibly even more useful than the EEG. Because the SQUID has better spatial resolution with the same temporal resolution, we can observe not only what the brain is doing from moment to moment but also more exactly which part of the brain is doing it. Chapter 17 reviews some early work on 1= f noise in MEG time series. Neural activity in the brain is also accompanied by changes in the ¯ow of blood to the neurons and to their surrounding supporting glial cells, a phenomenon ®rst noticed by Mosso in the late 1800s and turned into a general hypothesis by Roy and Sherrington in 1890 (see Posner and Raichle 1994). Not until the 1970s, however, did these early observations bear fruit, in the form of several imaging techniques that are now quite commonly used to provide precise information on exactly where in the brain neural activity occurs in response to various cognitive tasks. One of these, called ``positron-emission tomography'' (PET), involves introducing a weak radioactive tracer into the brain and then detecting the photons produced by collisions of the positrons emitted by the tracer in the

Dynamical Cognitive Neuroscience

299

blood with the electrons that are part of every brain tissue molecule. More of the tracer builds up in regions of high blood ¯ow, where the brain tissue is more active; the greater number of collisions between positrons and electrons produces more photons to be detected. In another technique, called ``functional magnetic resonance imaging'' (fMRI), active regions of the brain are imaged by surrounding the brain with an intense, ¯uctuating magnetic ®eld with which deoxygenated hemoglobin molecules in the brain resonate, producing a resultant inhomogeneity of the magnetic ®eld that can be detected. Regions of greater neural activity cause local vasodilation, bringing in increased amounts of oxygenated hemoglobin and relatively decreasing the amounts of deoxygenated hemoglobin. On the other hand, because they depend on blood ¯ow, and blood ¯ow is slower to respond to neural activity than are electrical or magnetic ®elds, PET and fMRI have somewhat poorer temporal resolution than the EEG and the MEG. Some newer imaging techniques, which detect the scattered remnants of weak laser beams injected directly into the brain (Gratton and Fabiani 1998) also show promise of combining high spatial and temporal resolution while not harming the brain. All these imaging techniques have the potential to produce a unique type of time seriesÐthe temporal unfolding of a three-dimensional map of neural activity in the brain. Ideally, the spatial scales would range from the microscopic to entire functional areas; the temporal scales, from milliseconds to minutes or even hours. Although techniques to analyze this enormous, complex mass of data are still in their infancy, such time series, obtained noninvasively while the brain engages in various cognitive tasks, indeed while similar time series of behavioral data are being collected, would allow the many links between behavior, cognition, and brain activity to be explored in detail. Some aspects of this relationship are already beginning to be studied, as discussed in section 33.2. 33.2

Brain Dynamics

Brains evolve (see chap. 32), and individual brains are born, develop into maturity, age, and ®nally die. Although the basic structure of the brain seems to be ®xed for each species, individual brains do not attain this structure invariably. Deprivation of experience at critical periods leads to structural abnormalities that persist, whereas exposure to experience leads a brain to develop more closely to its potential. Even the adult brain is plastic. For example, if the tactile cortex of an adult monkey is deprived of innervation from an arm, the part of the tactile cortex that previously

300

Chapter 33

encoded touch and pain information from that arm is invaded by innervation from neighboring regions of skin and subsequently encodes touch and pain information from those regions instead (Pons et al. 1991). In humans, extensive musical training over many years results in permanent changes in the brain, both in enhanced tactile representations of the ®ngers used in playing the violin (Elbert et al. 1995), and in enhanced auditory responses to piano tones (Pantev et al. 1998). Such plasticity occurs in the short term as well. For example, after listening for three hours to music from which a small band of frequencies around 1,000 Hz had been removed, people appeared to have fewer auditory cortical neurons devoted to coding sound in that frequency region. This tendency reversed between listening sessions, suggesting that the frequency tuning of neurons in the auditory cortex depends dynamically on the current soundscape (Pantev et al. 1999). Even short-term training on frequency discriminations in a particular frequency region can enhance the brain's responses to the relevant frequencies (Menning, Roberts, and Pantev 2000). These studies show more than simply changes in the brain as a consequence of learning. They demonstrate that the very structure and organization of the brain are dynamical (see also Ward 2001). The dynamics of the brain are nowhere more apparent than at the scale of the individual neurons and their connections to other neurons. At one time, these relationships were thought to be adequately modeled by the in¯uence of one neuron on another neuron's rate of generating action potentials (``®ring rate''). This in¯uence is relevant to ``rate coding'' by neurons, that is, expressing the presence of a trigger feature by an increase or decrease in ®ring rate. Recently, however, a di¨erent kind of in¯uence has been gaining prominence (see Singer 1993). In this view, the timing of a neuron's action potentials is in¯uenced by the timing of its inputs from other neurons. Synchronization of neural activity is thus considered to be important to the way the brain's state evolves in time and for the way in which it generates cognitive and behavioral states. Indeed, it amounts to a second possibility for neural coding, one especially useful for solving what is described as the ``binding problem''Ðthe coding of necessary relations between neural activity in anatomically separated but functionally related brain areas (see Singer 1999). Because relaxation oscillators are easily entrained by forcing inputs (see chap. 31), we might suspect that such oscillators would play a role in contemporary theories of synchronization mechanisms. They do. A striking example is the theory of Nunez (1995, 2000), according to which, neurons do act very much like relaxation oscillators, producing standing and traveling waves of neocortical activity that in¯uence global

Dynamical Cognitive Neuroscience

301

and local oscillations and cognitive and behavioral states, even consciousness (see chap. 35; Wright and Liley 1996). The theory begins with the observation that synaptic action density (the number of active synapses in a particular volume of neocortex), averaged over millions of neurons, is measured crudely by scalp EEG. The approach emphasizes coherences (squared correlation coe½cientsÐclosely related to synchronization) between paired locations of EEG recordings as estimates of the interactions of the neural networks found in these brain regions. The general principles of the theory are (1) brain state is described by physiologically based control parameters that depend on neuromodulatory action and that determine prominent EEG frequencies; (2) neural networks operate in a background of standing waves of synaptic actionÐ oscillatory changes in the number of active synapses; (3) excitatory and inhibitory synaptic action density and action potential density are treated as macroscopic ®eld variables; (4) cognitive and overt behavior depend on, among other things, the levels of these variables in various brain regions. The goal is to describe the brain system's input-output relationship in a general theory stated thus: ^ f …x; t†Š ˆ g…x; t†; D‰

…33:1†

^ is a mathematical where f …x; t† is the input, g…x; t† is the output, and D operator that represents the dynamical properties of the system. This operator is quite complicated (as might be expected) and also certainly nonlinear, which means that there is no easy way to decompose it into subterms that can be dealt with easily. Nunez (2000) points out, however, that it may be possible to use local linear approximations (see chap. 25) for certain brain states, or over certain parameter values, and thus begin to work with the theory. He does this in the context of EEG data in Nunez 1995 and 2000. Although there is not the space here to describe this complicated theory and its empirical plausibility in detail, it dovetails impressively with the ideas I have been trying to express in this book. The multiple scales of dynamics mentioned in the chapter epigraph from Nunez 2000 are collapsed for pragmatic reasons to two prominent scales of oscillations in synaptic action density and their interactions. Local oscillations arise from interactions between neurons in relatively small, functionally segregated areas of the brain, connected in neural networks through both excitatory and inhibitory connections. Global oscillations arise from interactions between such segregated areas, with longer communication delays caused by the longer axonal pathways, and possibly by the many intervening synapses, involved. Under some conditions, described in more detail by Nunez (1995, 2000), it is possible to derive simple

302

Chapter 33

Figure 33.2 Local and global oscillations. (From Nunez 2000. Reprinted by permission of Cambridge University Press.)

expressions for local and global oscillations and their interactions. In one such case, where the nonlinearity of the dynamics is small and the various global-local resonances do not interact, a set of ®ve second-order di¨erential equations can be obtained, each one like the following equation: d 2Z dZ ‡ ‰kn2 ÿ …1 ‡ e†Z 2 ŠZ ˆ 0; ÿ e…1 ÿ Z 2 † dt 2 dt

…33:2†

which closely resembles equation 31.3, the van der Pol relaxation oscillator: Z stands for the synaptic action density (like voltage in the van der Pol equation), e is the damping parameter, and ‰kn2 ÿ …1 ‡ e†Z 2 ŠZ is the restoring force term. If the nonlinearity is stronger, and the resonances are coupled, more complicated versions of equation 33.2 apply (right side not equal to zero). Numerical solutions to one set of such equations are shown in ®gure 33.2 for various local oscillation frequencies (q). Both

Dynamical Cognitive Neuroscience

303

local (faster) and global (slower) oscillations are apparent in ®gure 33.2. It has been suggested that gamma and theta frequencies of the EEG interact in this way to limit working memory (Lisman and Idiart 1995). In this theory, one ``rehearsal'' of all working memory ``slots'' takes place during every (global) theta cycle of about 6 Hz, rehearsal of each item taking place at the (local) gamma frequency of about 40 Hz, or one item every 25 msec, the maximum rate at which short-term memory can be ``scanned'' (Sternberg, 1966). Thus some 7 G 2 items can be retained in working, or short-term, memory (Miller 1956). In the service of con®rming their model, Lisman and Idiart (1995) published plots of embedded oscillations recorded from both human auditory cortex and from rat hippocampus that look remarkably like those of ®gure 33.2. 33.3

Hybrid Models

The brain has a de®nite structure, both anatomical and functional and in part modular (Swindale 1990, 1998); this structure is plastic at many scales of space and time. Although the prospect of describing the brain in a way that integrates anatomical and functional structure with dynamics is daunting, it is nonetheless beginning to be attempted. An excellent ®rst pass is Arbib, EÂrdi, and SzentaÂgothai 1997 (see also Arbib and EÂrdi 2000). Primitive, but illuminating, models can be suggested for cognitive tasks such as memory scanning, attentional orienting, or navigating space, in terms of which brain areas ``light up'' in a particular temporal sequence. One particularly in¯uential such model has been the ``disengage-shiftengage'' model of visual attention reorienting. In this model, disengaging of visual attention from a current location in visual space is mediated by activity in the posterior parietal cortex, shifting to a new location in visual space is mediated by activity in the superior colliculus, and reengaging attention at the new location is mediated by activity in the pulvinar nucleus of the thalamus (e.g., Posner and Raichle 1994). Because attentional orientation changes intermittently over time (oscillates?Ðsee Large and Jones 1999), its time evolution would consist of a series of episodes of shifting (the above sequence); these episodes of shifting would alternate with dwell times of various durations depending partly on the nature of the cognitive task being done (e.g., reading) and partly on the person's ``span of attention.'' Ultimately, it should be possible to coordinate more sophisticated models of brain activity with equally sophisticated models of cognitive operations to produce models that integrate brain and cognitive or behavioral dynamics in the way envisioned by Arbib and colleagues (Arbib, EÂrdi, and SzentaÂgothai 1997; Arbib and EÂrdi 2000).

This page intentionally left blank

Chapter 34 Dynamical Computation

No artifact devised by man is so convenient for this kind of functional description as the digital computer. ÐHerbert A. Simon, The Sciences of the Arti®cial If the bits of computers are someday scaled down to the size of individual atoms, quantum mechanical e¨ects may profoundly change the nature of computation itself. ÐD. P. DiVincenzo, ``Quantum Computation''

34.1

Numerical Methods

The digital computer is probably the most important weapon in the armamentarium of the dynamical cognitive scientist. Because dynamical equations are so di½cult to understand and solve analytically, the computer is the tool of choice for analyzing dynamical systems. I have already alluded to numerical methods implemented on a digital computer for doing Fourier analysis, for simulating chaos and stochastic processes, for solving ordinary di¨erential equations, and for understanding nonlinear di¨erential and di¨erence equations (usually involving iteration of a function; see chap. 23). Computers are modeling tools par excellence. Indeed, it is hard to see how anyone could comprehend complex systems of di¨erential or di¨erence equations without computer analysis. With computer analysis, we can often see how a complex model behaves under various parameter regimes and thus come to understand it. We can also use computer methods to obtain new predictions, often counterintuitive consequences of ``innocent'' assumptions. Of all of the techniques mentioned or suggested in this book, perhaps the most mundane and yet useful one for dynamical approaches is that of numerical integration of a di¨erential equation. Because numerical integration can be used to solve any di¨erential equation approximately for

306

Chapter 34

certain ranges of variable and parameter values, and because di¨erential equations form the foundation of dynamical modeling, it is de®nitely worth knowing about. Integrating a di¨erential equation produces a solution to it that expresses the relationship between a set of variables over some range of their values. Although a given di¨erential equation may express how the changes in the state variables of a system depend on some other variables, especially on time in the dynamical case, the actual evolution of the system in time is expressed by the solution to the equation, thus the need to integrate. Remember, in this case integration is not done to obtain a measure of a set (for example the area under a probability density function), but rather to ®nd a function relating the system's variables to time. Although there are many sources of computer algorithms for numerical integration, the one I have found most useful is Press et al. 1992, which contains descriptions of algorithms for this and many other numerical techniques of use to dynamical modeling (including random number and probability distribution generators), and which comes with an impressive array of software to implement these algorithms and with a website for news and problem reports. To get a taste of numerical integration without going into its complexities, let us begin with an ordinary di¨erential equation (ODE) of any order. This equation can be reduced to a set of ®rst-order ODEs by choosing new variables, as FitzHugh did when deriving his neuron model from the van der Pol relaxation oscillator (see chap. 31). We can apply numerical integration to each equation in this set, one at a time. Consider what is called an ``initial-value problem'' for a single ODE: given xs at a starting time ts , we want to know what value xf has at some ®nal time, tf , or what value xi has at some discrete list of time points ti between ts and tf , perhaps to plot the function between ts and tf . The basic idea of numerical integration is to construct the trajectory of the function x…t† piece by piece by adding a series of small increments to the initial value, xs , each one consisting of a step-size multiplied by a derivative. The simplest way to do this is called the ``Euler method.'' Often used for modeling neurons, this method is also valuable for insight into what is happening in the numerical integration process: it approximates the solution function by a series of straight lines (local linear approximation; see chap. 25), with each new point calculated according to x n‡1 ˆ x n ‡ Dt
f …tn ; x n †;

…34:1†

where Dt is the small increment in time, and f …tn ; x n † is the right-hand side of the di¨erential equation we are integrating. As a simple example,

Dynamical Computation

307

consider the ®rst-order ODE dx ˆ at; dt which you might remember as equation 8.1, whose solution, xˆa

t2 ; 2

was equation 8.2. Let us say we want to plot the function that yields equation 8.1, but we do not know the solution, equation 8.2 (we pretend it is too di½cult to solve analytically). First we rewrite the di¨erentials, dx and dt, in equation 8.1 as di¨erences, Dx and Dt, and multiply both sides by Dt, producing Dx ˆ Dt
at:

…34:2†

Now we can calculate the value of x at a point Dt in the future from the starting point, t0 , by adding the increment, Dx, speci®ed by equation 34.2 to the starting value of x, x 0 : x1 ˆ x 0 ‡ Dx ˆ x 0 ‡ Dt
at:

…34:3†

Clearly, equation 34.3 can be applied recursively, generating a series of values, xi , that can be plotted against the successive values of t, yielding a graph of the function (the integral solution of the di¨erential equation). Thus, in the absence of an analytical solution of a di¨erential equation, we can use this and similar numerical techniques to obtain numbers that are close to those taken by the solution: we can plot orbits of such equations in phase space, even if we cannot solve them analytically. Figure 34.1 shows an example of a series of points generated in this way from equation 34.3 with a ˆ 0:5 and x ˆ 0 at t ˆ 0. It also shows the analytical solution, equation 8.2, plotted in the same graph for comparison. The Euler method generates a reasonable approximation, although there is some errorÐin this case, all of the solution points are overestimated by a small amount. For equations that are less well behaved, the error can be large at some points, for example, at points of in¯ection. This error can be reduced by using a more sophisticated method such as the fourth-order Runge-Kutta, which creates several Euler steps, one at beginning of each time interval, twice at the midpoint and once at a trial endpoint and the ®nal function estimate is calculated based on these four derivatives, hence ``fourth-order Runge-Kutta'' (see Press et al. 1992). Another approach to dealing with intractable dynamical system descriptions is to begin with a potentially interesting di¨erence equation

308

Chapter 34

Figure 34.1 Euler approximation and solution to Equation 34.2.

and study its behavior by computer. Gregson (1988) studied a complex cubic di¨erence equation in this way as part of a model of psychophysical judgment. His di¨erence equation can be written as Yj‡1 ˆ ÿa…Yj ÿ 1†…Yj ‡ ie†…Yj ÿ ie†;

…34:4†

where the variable Y is complex, that is, each value of Y is a vector composed of a real value and an imaginary value, often written as y ˆ x ‡ iz, p  where the multiplier i ˆ ÿ1 creates the imaginary part. Through his computer analysis of this function, Gregson was able to ®nd parameter values for which chaos appeared in the imaginary component and a¨ected the time evolution of the real component. This is an excellent example of how computer exploration of a complicated, in this case nonlinear, difference equation can inform dynamical theory. 34.2

Neural Network Models

The computational power of the digital computer is critical to the implementation of neural network models, also called ``parallel distributed processing models'' and ``connectionist architectures,'' which have augmented traditional symbolic arti®cial intelligence in computer science (e.g., Hudson and Cohen 2000; Lewis, Jagannathan, and YesËildirek

Dynamical Computation

309

1999), and which have also proved capable of explaining many perceptual and cognitive behaviors in humans and other animals (e.g., Rumelhart and McClelland 1986; Parks, Levine, and Long 1998). Neural network models are inherently dynamicalÐthey must ``learn'' to model a particular behavior or to accomplish a particular task. Such learning is accomplished by implementing a graphlike structure of processing elements whose interconnections change over time until the structure is correctly accomplishing the task before it. Thus neural network models are potentially capable of realizing the program described in chapter 9 of building dynamical models on a plausible structural base. Any neural network model must have a minimum set of elements, each of which can be more or less complicated (e.g., Golden 1996; Rumelhart, Hinton, and McClelland 1986). Figure 34.2A shows these aspects in a schematic form. First, the network is composed of a set of processing units, ui , i ˆ 1 to N. In the simplest neural network in use today, there are three types of units: input, output, and hidden, as shown in ®gure 34.2B. Input units receive input from an environment and connect to either hidden units or to output units, whereas output units provide output signals to that environment and receive connections from either input or hidden units. Hidden units are called that because they do not interact directly with the environment: they receive inputs from either input or other hidden units and output to either output units or other hidden units. Networks without hidden units are called ``perceptrons,'' which lack the computational power to model the complexities of human cognition (Minsky and Papert 1967). Each processing unit has a state of activation, ai , and an output function, oi …t† ˆ fi …ai †, that transforms the unit's state of activation into an output signal. Processing units are connected to one another by weighted connections (similar to a directed, weighted, graph; see chap. 9) that together represent what the network ``knows'' and thus how it responds to any input. The connection weights, wi; j , are positive if the output from ui excites uj , and negative if ui inhibits uj . The size of wi; j represents the magnitude of the e¨ect of ui on u j . It is possible for a single unit to have many inputs, and these are referred to collectively as ``netj .'' The rule of propagation speci®es for each type of connection what the resulting net input is into each unit; it is often simply the weighted sum of the inputs of a given type. The activation rule speci®es how the various net inputs to a unit are combined with its current state of activation through a function, F, to produce its new state of activation for the next time period. The activation rule can be quite simple, such as the sum of net inputs and the current activation level, or it can be more complicated, such as a logistic or other threshold function or a Gaussian. The activation function

310

Chapter 34

Figure 34.2 (A) Elements of neural network models (From Rumelhart, Hinton, and McClelland 1986. Reprinted by permission of The MIT Press.); (B) three-layer neural network.

Dynamical Computation

311

might be some complicated processing that involves many computations, and that might even implement a semiautonomous agent. Initially, processing units are connected randomly or uniformly, with meaningless connection weights. The connection weights, or strengths, change over time as a result of implementation of a learning rule, as the network learns to do its task. Most learning rules are variants of the famous Hebbian learning rule (Hebb 1949), in which the connection between two units is strengthened if both units are highly active at the same time, and weakened if they are active at di¨erent times. Quite often the rule is modi®ed so that weights are changed in proportion to the di¨erence (delta) between the achieved activation and a target activation for the receiving unit; this is called the ``delta rule.'' In the backpropagation process, an error signal calculated at the output units is propagated backward into the network, so that the weights on each layer are changed successively based on the ``error'' calculated for that layer. Finally, in most neural network models, the environment is represented as some probability distribution over the set of possible input patterns, perhaps varying over time as well. Although neural network models are inherently dynamical, and are a foundation stone of modern arti®cial intelligence research, they have proven di½cult to accept as models of human cognition (e.g., Green 1998). Ironically, their strength is also their weakness in this regard: their ``knowledge'' of the problem domain is represented in the connection weights between the processing units. The pattern of activation of the hidden units, in particular, appears on the surface to be fairly random with respect to the critical aspects of the patterns presented to the input units. The problem is that the processing units are all the same, and often bear no direct relationship to the concepts of the phenomenon being modeled. This problem is common to all distributed representations (see Lashley 1950), in contrast to more traditional models, in which the model units themselves represent processing modules or conceptual units more directly related to the concepts of the problem domain. For example, a neural network model of visual pattern recognition may correctly classify patterns, even partial and novel patterns, but it is di½cult to see exactly how it is doing that task. In contrast, when a more traditional model performs, say, recognition by feature detection, it is clear what is going on because the concepts of the model are the same as those of the problem Ðin this case, features of visual patterns. Connectionist models must be analyzed to ®nd representations of cognitive concepts and to discover how the representations arose through the learning process, not an easy task because usually the more useful the model, the more complex it is. Progress is being made, however, on developing methods by which to

312

Chapter 34

analyze neural network models for representations of cognitive concepts. For example, Berkeley and Dawson (1995) reported ®nding clear representations of rules for making various logical deductions in the patterns of activations of the hidden units in their neural network. Others have used such multivariate techniques as factor analysis and discriminant function analysis to accomplish similar analyses. Consider a neural network made up of relaxation oscillators (see chap. 31). Here the processing done by individual units (activation rule) would be fairly complex, although the propagation rule could be simpler. Inputs from various units would represent stochastic and deterministic forcing, and di¨erent activation rules could endow di¨erent units with agentlike capabilities. Networks of similar processing units have been explored to some extent by Grossberg (1988), who introduced the concept of adaptive resonance in classifying patterns. In adaptive resonance, top-down signals from higher-level units interact with bottom-up signals from lower-level units to stabilize the network and create resonance when the input is suf®ciently close to a stored category prototype. During learning, both the top-down and bottom-up connection weights are changed adaptively, so that the category prototypes change to re¯ect experience. More purely mathematical work on systems of coupled oscillators can be found in Kopell 1988 and Rand, Cohen, and Holmes 1988. Another way to modify traditional neural networks is to introduce temporal coding, that is, to make the timing of signals between processing units important to their activation by those signals. This was done by Berger and Liaw (2000) in a simple neural network (11 neurons and 30 links) designed to do recognition of isolated spoken words presented in ``babble'' noise (essentially nonsense talking), a very di½cult task for humans. In this network, the learning rule allowed the temporal properties as well as the magnitudes of the connection weights to change as learning progressed. The network recognizes human speech in babble noise much better than any previous models can, including neural networks with thousands of neurons and connections, and also better than humans can, even when the noise is hundreds of times as intense as the speech signal. Thus there are two important timescales in neural networks: the scale of the learning and the scale of the signals between the units, which is tuned by the learning process. 34.3

Exotic Computers

Digital computers are composed of logic circuits implementing elementary (Boolean) logic: OR, NOT, AND, XOR, NAND, NOR, and so

Dynamical Computation

313

forth, and memory registers consisting of functionally organized groups of ¯ip-¯ops. A digital computer is a logic machine that simply repeats the same logic operations over and over again in di¨erent sequences. These sequences of operations, though speci®c to the task being accomplished, are implemented in the same way regardless of whether the task is numerical computation or human speech processing. Essentially, the various patterns of bits and the operations and groups of operations performed on them represent di¨erent symbols during di¨erent tasks. These symbols must be interpreted by a user. In a sense, the typographical level of operation, the level of simple logical operations performed on bits of information, is insulated from the semantic level of the user interpretation (Hofstadter 1979). Recently, large groups of interlinked digital computers have been formed to allow massively parallel processing. Some neural networks, such as that of Berger and Liaw (2000), are actually implemented as groups of simple computers, rather than being simulated by a single computer. Any digital computer, including a massively parallel one, can be emulated (the same calculations performed in the same way) by a universal Turing machine (Turing 1936) consisting of only an unbounded storage tape, a movable tape read/write head, and a set of elementary operations on the bits encoded on the tape. Such a machine has been said to be able to perform all physically possible computations (the Church-Turing thesis), although this is di½cult to prove. Indeed, there have been proposals of systems that could compute things that the Turing machine cannot. One such is the analog shift map of Siegelmann (1995), created from a chaotic dynamical system similar to the logistic di¨erence equation discussed in chapter 24. An ``analog shift'' replaces one dotted substring (consisting of symbols and one dot, for example, 0.1) with another equally long one according to a function G; then the string is shifted an integer number of places to the left or to the right (relative to the dot) according to another function, F. For example, starting with the sequence 1010.101, if the function G operates on the subsequence 0.1 by replacing ``0.1'' with ``1.1'', then G…1010:101† ˆ 1011:101. If F then speci®es that the dot is to be moved one place to the left, then F …1011:101† ˆ 101:1101. This process continues until the functions G and F specify no further changes; the result of the computation is the ®nal sequence, and the input-output map is the transformation from the initial dotted sequence to the subsequence to the right of the dot in the ®nal sequence. Siegelmann (1995) proved that this process can compute a class of functions that contains the class of functions computed by the universal Turing machine as a subset, thus it represents a more powerful computer. Moreover, the analog shift map

314

Chapter 34

computes in a way similar to that of neural networks and analog computers, and could possibly describe real physical systems. Realizations of this hypothetical computer could come in the form of physical systems that behave chaotically, making it possible to exploit chaos as a computational device, and also making it possible to simulate very complicated neural networks directly. Among the other exotic, hypothetical computers that have been suggested in recent years (including ones based on DNA), at least one has properties that might become useful for dynamical cognitive science. This is the quantum computer (e.g., Feynman 1985; DiVincenzo 1995). The most important di¨erence between classical digital computers and quantum computers is in the state of the elements that encode the user's symbols. A digital computer's ¯ip-¯ops always have a de®nite state at any moment of time, for example, an 8-bit register might read 10101111. In a quantum computer, however, the bits of information are sometimes described by a wave function (see chap. 13). When this is the case, as during a computation, the di¨erent possible states of a group of bits are superimposed in the wave function C ˆ aj10101111i ‡ bj11111111i ‡ cj00000000i ‡


;

…34:5†

where a; b; c; . . . are complex numbers whose phases (the parts multiplied by i) can be such that the di¨erent possible states interfere either constructively or destructively. When the state of the computer is measured, this wave function collapses and the bits assume one of the possible states with probabilities jaj 2 ; jbj 2 ; jcj 2 ; . . . The trick is to make the incorrect answers to a computation interfere destructively so that their probabilities are vanishingly small, whereas that of the state corresponding to the correct answer is near one. If this can be done, because complex calculations are accomplished in parallel across huge numbers of bits, they can be accomplished in much less time than in a conventional digital computer. With this much computing power, it should prove possible to emulate the human brain's activity over time at any necessary level of detail (cf. Tipler 1994). Moreover, it has even been suggested that the human brain implements a quantum computer that is responsible in some way for the noncomputable aspects of human consciousness (Hamero¨ 1998; Penrose 1994; but see also Tegmark 2000). Thus understanding quantum computers could provide insights into the temporal unfolding of human cognition more directly than attained by simulations of complicated dynamical models.

Chapter 35 Dynamical Consciousness

O, what a world of unseen visions and heard silences, this insubstantial country of the mind! . . . And the privacy of it all! . . . This consciousness that is myself of selves, that is everything, and yet nothing at allÐwhat is it? And where did it come from? And why? ÐJulian Jaynes, The Origin of Consciousness in the Breakdown of the Bicameral Mind Consciousness, then, does not appear to itself chopped up in bits. . . . It is nothing jointed; it ¯ows. . . . In talking of it hereafter, let us call it the stream of thought. . . . ÐWilliam James, The Principles of Psychology We propose that a large cluster of neuronal groups that together constitute, on a time scale of hundreds of milliseconds, a uni®ed neural process of high complexity be termed the ``dynamic core,'' in order to emphasize both its integration and its constantly changing patterns. ÐGiulio Tononi and Gerald M. Edelman, ``Consciousness and Complexity'' Human consciousness is itself a huge complex of memes (or more exactly memee¨ects in brains) that can best be understood as the operation of a ``von Neumanesque'' virtual machine implemented in the parallel architecture of a brain that was not designed for any such activities. ÐDaniel C. Dennett, Consciousness Explained

35.1

Consciousness

Arguably the greatest challenge of dynamical cognitive science in the future will be to provide a basis for understanding the unique and fascinating phenomena of human consciousness. The project is of gigantic proportions, and although breakthroughs may come in the next few years, indeed, may have already been initiated, it is unlikely that a detailed understanding of the mechanisms of human (and other animal) consciousness will be attained in less than many years (the title of Dennett

316

Chapter 35

1991 to the contrary notwithstanding). Although consciousness has been neglected for years in psychology, the group studying it has become both vast and interdisciplinary, with physicists and neuroscientists joining the psychologists and philosophers in telling their stories around the camp®re. Here I will touch on a few of the most promising (in my opinion) and dynamically relevant ideas that swirl around this topic. The consciousness so aptly and poetically alluded to in the ®rst two chapter epigraphs clearly includes wakefulness, our awareness of or attention to environmental events, our ability to report on our mental states, our ability to remember episodes in our past life at will, our ability to contemplate our own mental states and our own behavior, and, above all, our experience of mental life, what the philosophers call ``what it is like to be me (or you, or a bat).'' Chalmers (1996) argued that explaining the ®rst several items on this list (and some others) is ``easy'' relative to explaining the last one, that of experience, which he called the ``hard problem.'' Indeed, we could not hope to use the ordinary methods of cognitive science (including dynamical cognitive science) to achieve a satisfactory account of the problem of experience because nothing in any functional or mechanistic explanation of any aspect of consciousness could possibly entail that the function or mechanism in question would be accompanied by experience. Chalmers (1996) speculated that the only way to deal with experience is to grant it fundamental status, in the same way that we grant electrons, photons, and quarks fundamental status. We could do this by asserting that at least some information (the di¨erence that makes a difference; Bateson 1972) has two distinct aspects or properties, one physical and the other phenomenal. This is consistent with suggestions that information, rather than matter, is the most fundamental physical concept (e.g., Wheeler 1990), and that information can take di¨erent aspects depending on the point of view of the observer (Pribram 1986). We can study the phenomenal aspect, as we study the physical one, but there is no use asking why there is a phenomenal aspect to information, any more than it makes sense to ask why there is a physical aspect. This view denies that conscious experience is an eternal mystery (traditional dualism) but also denies that it will be possible to explain it in a reductionist functional or mechanistic way (scienti®c materialism, a form of monism). Dennett (1991) has argued compellingly that we should abandon the traditional view of the mind/brain as having a place, the Cartesian theater, where ``it all comes together,'' where consciousness lives. Instead, he has proposed, the mind/brain contains many parallel channels of information processing (he called them ``multiple drafts'') and that di¨erent parts of

Dynamical Consciousness

317

this stream of parallel processing form the dominant narrative, a virtual von Neuman serial computer, at di¨erent times. There is no Cartesian theaterÐthe theater of consciousness is the entire brain. Indeed, from the evidence gathered thus far, it seems clear that some kind of massive parallelism must be the case, and that it must form a foundation stone of any explanation of how the brain generates experience and the other attributes of consciousness. Why are we conscious? Many (e.g., Shallice 1972) have sought to discover something that consciousness does, that can be accomplished by no other mechanism. One prominent recent suggestion (e.g., Delacour 1995; Donald 1995) is that the usefulness of the huge computational power of the massively parallel organization of the human brain (and mammalian brains in general) is somewhat compromised by a lack of integration; at any moment each parallel process could be doing its own thing, and the whole organism could be paralyzed or could oscillate wildly between inconsistent perceptions and actions. Consciousness, by allowing competing processes to interact in the same ``global workspace'' (e.g., Baars 1988), could achieve a level of integration that would counteract these negative qualities of parallel computing. This ``virtual serial von Neumanesque'' aspect of the mind/brain would be responsible for intentionality and models of the self in the worldÐor, at least, these would be the metaphors used by modern humans to describe this integrative activity (cf. Jaynes 1976). What form could the brain activity associated with consciousness take? The modern search for the neural correlates of consciousness was greatly stimulated by Crick and Koch 1990. One neural correlate of consciousness they suggested, might be found in the roughly 40 Hz oscillations discovered by Gray and Singer (1989) in the cat visual cortex (see also Singer 1999), where neural activity was synchronized via locking to a common ®ring frequency. Not only could such synchronization bind together the disparately coded features (color, shape, motion, etc.) of visual stimuli; perhaps it could also be the sine qua non of consciousness (or at least visual awareness) itself. Whereas Crick and Koch continued to search for some property of individual neurons that would confer consciousness, several others focused on the temporal synchrony itself as the critical feature. One of the most impressive demonstrations of the importance of such synchronization builds on an ingenious technique for studying consciousness ®rst exploited by Logothetis and Schall (1989): presenting visual stimuli that engage in binocular rivalry (cf. Levelt 1968) and simultaneously directly recording neural activity from neurons in

318

Chapter 35

visual cortex that are responding to these stimuli. In binocular rivalry, con¯icting stimuli, such as a square-wave grating with stripes oriented vertically and another with stripes oriented horizontally, are presented at the same time, one to each eye (see ®g. 35.1). The viewer is generally only aware of (conscious of ) one of the two stimuli, and the ``seen'' grating alternates randomly. What happens in the visual cortex when the ``seen'' grating switches from one to the other, that is, when awareness changes? According to one hypothesis, the ®ring rates of various neurons undergo modulation as the view changes, with the neurons responding to the newly ``seen'' grating increasing their ®ring rate just as the view changed, and those responding to the newly ``unseen'' grating decreasing their ®ring rate at the same time. Although such modulation does not consistently occur in the cortical areas that do the early processing of visual stimuli (areas V1, V2, and V3), it does appear to occur in inferotemporal cortex, where the later stages of visual processing take place (Sheinberg and Logothetis 1997) According to a second hypothesis, in addition to, or instead of, modulation of ®ring rate, awareness changes are accompanied by a change in the synchrony of the ®ring of the various neurons that encode the stimuli. This hypothesis was con®rmed for neurons in areas V1 and V2 in a clever study of binocular rivalry in cats by Fries et al. (1997). Figure 35.1 shows their experimental setup and results. The cat was watching mirrors that presented one drifting (slowly moving in the direction of the arrow) grating to each eye (panel A). Under rivalry conditions, with two gratings present, which one the cat was aware of was inferred from which one was causing optokinetic nystagmus (slow tracking eye movements interrupted by abrupt ``return'' eye movements that indicate that the cat was tracking the grating). Because the eyes move together, the directions of the slow and abrupt eye movements indicated which grating was being ``seen'' at any moment. When only one grating was presented to only one eye as a control (panels B and D), the neurons responding to that grating were quite synchronized (large amplitude and oscillatory nature of correlation functions below the gratings in the ®gure) and there was no di¨erence in the synchrony between the two gratings. When both stimuli were presented together, however, the neurons responding to the ``seen'' grating (panel C) were much more highly synchronized than were those responding to the ``unseen'' grating (panel D). These are dramatic results, but do they apply to human consciousness? Much recent work using electroencephalography and magnetoencephalography (EEG and MEG; e.g., Sauve 1999, see chap. 33) suggests that

Figure 35.1 Neural synchronization under binocular rivalry. (From Engel et al. 1999.)

320

Chapter 35

they do. One result using MEG, in particular, is especially suggestive. Tononi et al. (1998) reported that, when humans are conscious of a particular stimulus among rivaling visual grating stimuli, there is widespread coherence between MEG measurements at various brain sites and the MEG measurements of the activity of the neurons in visual cortex responding to the seen grating, whereas such coherence is lacking for the unseen grating (coherence relates to the square of the correlation coe½cient between two time series). Tononi et al.'s experimental setup was similar to that of Fries et al. (1997), except that the rivaling gratings were ¯ickered at di¨erent temporal frequencies. Because visual cortical neurons phase lock their ®ring to the ¯icker frequency of stimuli to which they are responding, the power in the appropriate frequency band of a Fourier spectrum of the MEG time series could be compared across brain sites when a given stimulus was ``seen'' and when it was suppressed. These comparisons indicated that widespread, coherent neural ®ring was associated with consciousness of a particular grating. Not all neural synchrony is associated with consciousness, however. The rigid, hypersynchronous ®ring characteristic of petit mal seizures is associated with an unconscious state. Apparently, before synchrony becomes associated with consciousness, there must be a large number of possible patterns of neural synchrony, of which the occurring pattern is only one (di¨erentiation). Tononi and Edelman (1998) suggested that both di¨erentiation and integration are fundamental aspects of consciousness, and that both are achieved by the emergence of a dynamic core, namely, ``a large cluster of neuronal groups that together constitute, on a time scale of hundreds of milliseconds, a uni®ed neural process of high complexity'' (p. 1849). The dynamic core is similar to Kinsbourne's ``dominant focus'' (1988) and to Dennett's von Neumanesque computer (1991), in both of which massively parallel computing in the brain is integrated in consciousness. Various aspects of the current sensory and cognitive activity constitute the dynamic core and are thus ``in consciousness,'' whereas other aspects, being outside the dominant focus, are outside of consciousness. Neural synchrony across widespread areas of the brain is the ``force'' that binds various neural activity into the dynamic core, with unsynchronized activity continuing outside of awareness (with often signi®cant consequences, as ideas ``pop'' into our minds, problem solutions ``just appear,'' and our bodies react in unpredicted ways to unperceived stimuli). An interesting implication of this view, consistent with Dennett's multiple drafts model (1991), is that time and space are ``smeared'' at the scale of the dynamic core process: several hundreds of milliseconds across the spatial extent of

Dynamical Consciousness

321

the brain. Because consciousness itself is a temporally smeared process it makes no sense to ask exactly when some aspect of the brain's information processing becomes conscious relative to when a stimulus is presented in eotemporal time (see Dennett and Kinsbourne 1992 for a deep discussion of this implication). The dynamic core generates its own biotemporal or nootemporal scales (see chap. 4), according to which conscious events (measured in clock time) are temporally indeterminate. The plausibility of this view has been demonstrated by a large-scale simulation of thalamocortical circuits in the brain (65,000 neurons and 5 million connections; Lumer, Edelman, and Tononi 1997). In this simulation, individual neurons were modeled as integrate-and-®re neurons (see chaps. 22 and 31) described by tm

X dVi …t† ˆ ÿVi ‡ E0 ÿ gj …t†…Vi ÿ Ej †; dt j

…35:1†

where Vi …t† is the voltage across the cell membrane at time t, E0 is the resting potential, gj …t† is the synaptic potential generated by a spike at synapse j at time t (realistically modeled as several di¨erent types of synapses), Ej is the reversal potential for synapse j, and tm is a passive membrane time constant set at 16 msec for excitatory and 8 msec for inhibitory neurons. Whenever the membrane potential reached a threshold (about ÿ51 mV) it was reset to ÿ90 mV and a spike was recorded. As illustrated in ®gure 35.2, neurons were grouped into various layers, with various pathways and reentrant loops. Based on other evidence, the reentrant loops are particularly important to visual consciousness (Lamme and Roelfsema 2000), probably because they facilitate acrossarea synchrony. An attempt was made to realistically represent the various pathways in the visual system, even to the inclusion of primary (subscript p in ®g. 35.2) and secondary (subscript s) areas in each region of the brain, and also several cortical layers. Visual inputs consisted of two superimposed square-wave gratings, with added noise, moving in perpendicular directions. When this neural network was simulated with continuous input, multilevel synchronous oscillations at roughly 50 Hz emerged, as shown in ®gure 35.3 by the population-averaged membrane potentials from cells in various layers of the model (based on from several hundred to several thousand simulated neurons in each layer). These oscillations were not programmed into the simulation in any way; they simply occurred as a consequence of the interactions between the neural activity in the various simulated brain areas in response to the stimulus input. Variation of parameters such as synaptic strengths, inhibitory time

322

Chapter 35

Figure 35.2 Organization of neural systems that generates spontaneous oscillations. (From Lumer, Edelman, and Tononi 1997. Reprinted by permission of Oxford University Press.)

Figure 35.3 Spontaneous oscillations from a simulated thalamocortical system. (From Lumer, Edelman, and Tononi 1997. Reprinted by permission of Oxford University Press.)

Dynamical Consciousness

323

constants, transmission delays, and lesions of various parts of the network a¨ected the frequencies, but not the presence, of oscillations. How does the peculiarly human feel to consciousness arise? We humans have a sense of self that often dominates our conscious awareness. We are aware of being aware and we talk about being aware, even to ourselves. The awareness of self seems to be restricted to humans, apes (chimpanzees, gorillas, and orangutans), and whales (e.g., dolphins and killer whales). Such awareness can be ascertained by testing whether an animal learns to recognize itself in a mirror (Gallup 1977), and to reverse roles in a social situation (Povinelli 1993). There is clearly a basis for the construction of a nonverbal self-concept in the basic facts of perception and social interaction; for example certain objects (our hands, our feet) cannot get smaller than a certain apparent size, no matter how we move about, whereas others (our car, a tree) can be made inde®nitely small simply by moving in a particular direction (see Bermudez 1998 for a detailed discussion). Without language, animals can only have such a nonverbal selfconcept based on perceptual-motor schemata, similar to a preverbal child, whereas with language, things change dramatically. Children learn to have a verbal self-concept, so that they can consistently use the word ``I'' to refer to behaviors, desires, and states of a particular entity (to which they have privileged access), and ``you,'' ``he,'' ``she,'' and ``they'' to refer to the actions, desires, and (inferred) states of everyone else. They appear to do this via a massive infusion of memes (Dennett 1991), simple idea units introduced by Dawkins (1976) to describe how cultural information seems to undergo selective evolution similar to that experienced by genetic information. (Cavelli-Sforza 1981 expands on this theme with a quantitative model of cultural transmission based on standard evolutionary theory.) Dennett (1991) probably was wrong, however, to assert that these memes created from scratch the von Neumanesque serial processor running in the parallel architecture of the brain. Apparently, the complexly interlinked relaxation oscillators that make up the brain spontaneously self-organize into the dynamic core. Without language and the memes it makes possible, however, even a human would be left with a simple, nonverbal, perceptual-motor concept of self, and could not achieve elaborate simulations of self in the metaphorical world to which verbal humans are prone (see Jaynes 1976; Kahneman and Tversky 1982). All of this has not yet answered the ``hard problem.'' Why is the synchronized neural processing of the dynamic core accompanied by ``experience?'' What is it about spatially distributed, temporally synchronized neural processing that generates experience? It is possible that exotic

324

Chapter 35

mechanisms are responsible. Hamero¨ (1998) and Penrose (1994) have suggested that the brain acts as a quantum computer (see chap. 34) and that this action is responsible for consciousness: physically, conscious awareness is a quantum ®eld e¨ect, similar to an electromagnetic ®eld. Hamero¨ (1998) suggested that anesthetics interrupt the necessary physical conditions necessary for the large-scale quantum coherence that generates consciousness, perhaps by binding electrons in the microtubule skeletons of the neurons so that they cannot participate in the quantum ®eld (called a ``Bose-Einstein condensate''). Although it has been argued that the timescale for such a ®eld in the brain is much too short to approximate the time scale of consciousness (Tegmark 2000), the dynamic core of neural processing may indeed generate some kind of ®eld e¨ect, possibly electromagnetic, that results in ``subjectivity.'' Certainly electromagnetic ®elds are what is measured by MEG and EEG, and these are more closely associated with cognitive states than are the detailed activities of individual neurons (e.g., Nunez 2000). These assertions must remain speculations, but they refer to provocative questions that will stimulate dynamical cognitive science for many years, even though we may never answer them (see Kurthen, Grunwald, and Elger 1998). 35.2

Unity of Science

It is appropriate to close this book with a discussion of consciousness, the study of which is inherently dynamical and interdisciplinary, with a strong need for complementary explanations. As we pass into an era when machines may think, when intelligence may evolve beyond that of humans (see Fogel 2000; Pfeifer and Scheier 1999) and quantum computation may allow us to emulate the human brain at the level of quarks (Tipler 1994), dynamical cognitive science, informed by complementary approaches from all of the sciences, will play a central role in our continuing e¨ort to understand ourselves and our place in the cosmos.

References

Alvarez de Lorenzana, J. M. (2000). Closure, open systems and the modeling imperative. In J. R. Chandler and G. Van der Vijver, eds., Closure, Emergent Organizations and Their Dynamics, 91±99. New York: New York Academy of Sciences. Alvarez de Lorenzana, J. M., and Ward, L. M. (1987). On evolutionary systems. Behavioral Science, 32, 19±33. Anderson, J. R. (1976). Language, Memory, and Thought. Hillsdale, NJ: Erlbaum. Anderson, J. R., Matessa, M., and Leabiere, C. (1997). Act-R: A theory of higher level cognition and its relation to visual attention. Human-Computer Interaction, 12, 439±462. Arbib, M., and EÂrdi, P. (2000). PreÂcis of Neural Organization: Structure, Function and Dynamics. Behavioral and Brain Sciences, 23, 513±571. Arbib, M., EÂrdi, P., and SzantaÂgothai, J. (1997). Neural Organization: Structure, Function, and Dynamics. Cambridge, MA: MIT Press. Ashby, W. R. (1958). Requisite variety and its applications for the control of complex systems. Cybernetics, 1, 83±99. Baars, B. J. (1988). A Cognitive Theory of Consciousness. Cambridge: Cambridge University Press. Baggott, J. E. (1992). The Meaning of Quantum Theory. New York: Oxford University Press. Baird, J. C. (1997). Sensation and Judgment: Complementarity Theory of Psychophysics. Mahwah, NJ: Erlbaum. Bak, P. (1990). Self-organized criticality. Physica A, 163, 403±409. Bak, P., Tang, C., and Weisenfeld, K. (1987). Self-organized criticality: An explanation of 1=f noise. Physical Review Letters, 59, 381±384. Bateson, G. (1972). Steps to an Ecology of Mind. New York: Ballentine. Beckers, R., Deneubourg, J. L., Goss, S., and Pasteels, J. M. (1990). Collective decision making through food recruitment. Insectes Sociaux, 37, 258±267. Beer, S. (1974). Designing Freedom. Toronto: Canadian Broadcasting Company Publishing Corporation.

326

References

Beer, R. D. (2000). Dynamical approaches to cognitive science. Trends in Cognitive Sciences, 4, 91±99. Bell, D. A. (1960). Electrical Noise. London: Van Nostrand. Benzi, R., Sutera, S., and Vulpiani, A. (1987). The mechanism of stochastic resonance. Journal of Physics, A14, L453±L457. Berge, G. (1985). Graphs. Amsterdam: North-Holland. Berger, T. W., and Liaw, J. S. (2000). Machine demonstrates superhuman speech recognition abilities. USC News Service [On-line]. Available: hhttp:// www.usc.edu/ext-relations/news_service/releases/stories/3013htmli. Berkeley, I. S. N., and Dawson, M. R. W. (1995). Density plots of hidden value unit activations reveal interpretable bands. Connection Science, 7, 167±187. Bermudez, J. L. (1988). The Paradox of Self-Consciousness. Cambridge, MA: MIT Press. Bezrukov, S. M., and Vodyanoy, I. (1997). Stochastic resonance in non-dynamical systems without response thresholds. Nature, 385, 319±321. Boon, J. P., and Decroly, O. (1995). Dynamical systems theory for musical dynamics. Chaos, 5, 501±508. Born, M. (1962). Einstein's Theory of Relativity. New York: Dover. Bouisson, M. (1960). Magic: Its Rites and History. London: Rider. Box, G. E. P., and Jenkins, G. M. (1976). Time Series Analysis: Forecasting and Control. Rev. ed. San Francisco: Holden-Day. Braida, L. D., and Durlach, N. I. (1988). Peripheral and central factors in intensity perception. In G. M. Edelman, W. Gall, and W. M. Cowan, eds., Auditory Function, 559±583. New York: Wiley. Bremermann, H. J. (1962). Optimization through evolution and recombination. In M. C. Yovits et al., eds., Self-Organizing Systems, 93±106. Washington, DC: Spartan. Bringhurst, R. (1995). The Calling: Selected Poems, 1970±1985. Toronto: McClelland and Stewart. Brock, W. A., Dechert, W. D., and Scheinkman, J. (1986). A test for independence based on the correlation dimension. Department of Economics, University of Wisconsin, Madison, University of Houston, and University of Chicago. Brock, W. A., Hsieh, D., and LeBaron, B. (1991). Nonlinear Dynamics, Chaos, and Instability: Statistical Theory and Economic Evidence. Cambridge, MA: MIT Press. Brockwell, P. J., and Davis, R. A. (1996). An Introduction to Time Series and Forecasting. New York: Springer. Bryden, M. P. (1967). A model for the sequential organization of behaviour. Canadian Journal of Psychology, 21, 37±56. Budescu, D. V. (1987). A Markov model for generation of random binary sequences. Journal of Experimental Psychology: Human Perception and Performance, 13, 25±39.

References

327

Burgess, A. (1962). A Clockwork Orange. London: Heinemann. Bush, R. R., and Mosteller, F. (1955). Stochastic Models for Learning. New York: Wiley. Buss, D. M. (1999). Evolutionary Psychology: The New Science of the Mind. Boston: Allyn and Bacon. Casdagli, M. (1991). Chaos and deterministic versus stochastic non-linear modelling. Journal of the Royal Statistical Society B, 54, 303±328. Cavalli-Sforza, L. L., and Feldman, M. W. (1981). Cultural Transmission and Evolution: A Quantitive Approach. Princeton, NJ: Princeton University Press. Chalmers, D. J. (1996). The Conscious Mind: In Search of a Fundamental Theory. New York: Oxford University Press. Chan, K. S., and Tong, H. (2001). Chaos from a Statistical Perspective. New York: Springer. Clancy, W. J. (1997). Situated Cognition: On Human Knowledge and Computer Representations. New York: Cambridge University Press. Clancy, W. J. (1999). Conceptual Coordination: How the Mind Orders Experience in Time. Mahwah, NJ: Lawrence Erlbaum Associates. Clayton, K., and Frey, B. (1997). Studies of mental ``noise.'' Nonlinear Dynamics, Psychology, and Life Sciences, 1, 173±180. Collins, J. J., Chow, C. C., Capela, A. C., and Imho¨, T. T. (1996). Aperiodic stochastic resonance. Physical Review E, 54, 5575±5584. Collins, J. J., Imho¨, T. T., and Grigg, P. (1996a). Noise-enhanced information transmission in rat SA1 cutaneous mechanoreceptors via aperiodic stochastic resonance. Journal of Neurophysiology, 76, 642±645. Collins, J. J., Imho¨, T. T., and Grigg, P. (1996b). Noise-enhanced tactile sensation. Nature, 383, 770. Cordo, P., Inglis, J. T., Verschueron, S., Collins, J. J., Merfeld, D. M., Rosenblum, S., and Moss, F. (1996). Noise in human muscle spindles. Nature, 383, 769±770. Coren, S., Ward L. M., and Enns, J. T. (1999). Sensation and Perception. 5th ed. Fort Worth: Harcourt Brace. Corsini, G., and Saletti, R. (1987). Design of a digital 1=f v noise simulator. In C. M. Van Vliet, ed., Ninth International Conference on Noise in Physical Systems, 82±86. Singapore: World Scienti®c. Cowell, R. G., Philip, D. A., Lauritzen, A. L., and Spiegelhalter, D. J. (1999). Probabilistic Networks and Expert Systems. New York: Springer. Crick, F. H. C., and Koch, C. (1990). Towards a neurobiological theory of consciousness. Seminars in the Neurosciences, 2, 263±275. Crovitz, F. (1970). Galton's Walk : Methods for the Analysis of Thinking, Intelligence and Creativity. New York: Harper and Row. Cutler, C. D. (1997). A general approach to predictive and fractal scaling dimensions in discrete-index time series. Fields Institute Communications, 11, 29±48.

328

References

Darwin, C. (1871). The Descent of Man, and Selection in Relation to Sex. London: Murray Dawkins, R. (1976). The Sel®sh Gene. New York: Oxford University Press. Dawkins, R. (1986). The Blind Watchmaker. New York: Norton. De Bono, E. (1970). Lateral Thinking: Creativity Step by Step. New York: Harper and Row. Delacour, J. (1995). An introduction to the biology of consciousness. Neuropsychologia, 33, 1061±1074. Dement, W. C. (1974). Some Must Watch While Some Must Sleep. San Francisco: Freeman. Dennett, D. C. (1991). Consciousness Explained. Toronto: Little, Brown. Dennett, D. C., and Kinsbourne, M. (1992). Time and the observer: The where and when of consciousness in the brain. Behavioral and Brian Sciences, 15, 183± 247. Devaney, R. L. (1989). An Introduction to Chaotic Dynamical Systems, Second Edition. New York: Addison-Wesley. DeCarlo, L. T., and Cross, D. V. (1990). Sequential e¨ects in magnitude scaling: Models and theory. Journal of Experimental Psychology: General, 119, 375±396. Dirac, P. A. M. (1958). The Principles of Quantum Mechanics. 4th ed. Oxford: Clarendon Press. DiVincenzo, D. P. (1995). Quantum computation. Science, 270, 255±261. Donald, M. (1995). The neurobiology of human consciousness: An evolutionary approach. Neuropsychologia, 33, 1087±1102. Douglass, J. K., Wilkens, L., Pantazelou, E., and Moss, F. (1993). Noise enhancement of the information transfer in cray®sh mechanoreceptors by stochastic resonance. Nature, 365, 337±340. Dukas, R. (1998a). Introduction. In Dukas, ed., Cognitive Ecology: The Evolutionary Ecology of Information Processing and Decision Making, 1±19. Chicago: University of Chicago Press. Dukas, R. (1998b). Constraints on information processsing and their e¨ects on behavior. In Dukas, ed., Cognitive Ecology: The Evolutionary Ecology of Information Processing and Decision Making, 89±127. Chicago: University of Chicago Press. Dukas, R., ed. (1998c). Cognitive Ecology: The Evolutionary Ecology of Information Processing and Decision Making. Chicago: University of Chicago Press. Dukas, R., and Clark, C. W. (1995). Sustained vigilance and animal performance. Animal Behaviour, 49, 1259±1267. Dunne, J. W. (1934). The Serial Universe. London: Faber and Faber. Eccles, J. C. (1964). The Physiology of Synapses. Berlin: Springer. Einstein, A., H. Born, and M. Born (1969). Brief zum Max Born, 4 December 1926 in Einstein und Born Briefwechsel: 1916±1955. Kommentiert von Max Born. MuÈnchen.

References

329

Einstein, A. (1906). Zur Theorie der brownschen Bewegung. Annalen der Physik, 19, 371±381. Einstein, A. (1961). Relativity. The Special and General Theory. A Popular Exposition. New York: Crown Publishers. Elbert, T., Pantev, C., Weinbruch, C., Rockstroh, B., and Taub, E. (1995). Increased cortical representation of the left hand in string players. Science, 270, 305±307. Engel, A. K., Fries, P., KoÈnig, P., Brecht, M., and Singer, W. (1999). Temporal binding, binocular rivalry, and consciousness. Consciousness and Cognition, 8, 128±151. Enns, J. T., and Di Lollo, V. (2000). What's new in visual masking? Trends in Cognitive Sciences, 4, 345±352. Ericsson, K. A., and Simon, H. A. (1993). Protocol Analysis: Verbal Reports as Data. Cambridge, MA: MIT Press. Falmagne, J.-C. (1985). Elements of Psychophysical Theory. New York: Oxford University Press. Farmer, J. D., and Sidorowich, J. J. (1987). Predicting chaotic time series. Physical Review Letters, 59, 845±848. Faust, D., and Ziskin, J. (1988). The expert witness in psychological and psychiatry. Science, 241, 31±35. Fechner, G. T. (1860). Elemente der Psychophysik. Leipzig: Breitkopf and HaÈrtel. Feller, W. (1968). An Introduction to Probability Theory and Its Applications. Vol. 1. 3d ed. New York: Wiley. Feller, W. (1971). An Introduction to Probability Theory and Its Applications. Vol. 2. 2d ed. New York: Wiley. Feynman, R. P. (1965). Character of Physical Law. Cambridge, MA: MIT Press. Feynman, R. P. (1985). QED: The Strange Theory of Light and Matter. Princeton, NJ: Princeton University Press. Feynman, R. P., Leighton, R. B., and Sands, M. (1963). The Feynman Lectures on Physics. Reading, MA: Addison-Wesley. FitzHugh, R. A. (1960). Thresholds and plateaus in the Hodgkin-Huxley nerve equations. Journal of General Physiology, 43, 867±896. FitzHugh, R. A. (1961). Impulses and physiological states in theoretical models of nerve membrane. Biophysical Journal, 1, 445±466. Fogel, D. B. (2000). Evoluntionary Computation: Toward a New Philosophy of Machine Intelligence. New York: Institute of Electrical and Electronics Engineers. Fraser, J. T. (1982). The Genesis and Evolution of Time: A Critique on Interpretation in Physics. Amherst: University of Massachusetts Press. Freeman, W. J. (1991). The physiology of perception. Scienti®c American, 264(2), 78±85. Friedman, M. P., and Carterette, E. C. (1964). Detection of Markovian sequences of signals. Journal of the Acoustical Society of America, 36, 2334±2339.

330

References

Fries, P., Roelfsema, P. R., Engel, A. K., KoÈnig, P., and Singer, W. (1997). Synchronization of oscillatory responses in visual cortex correlates with perception in interocular rivalry. Proceedings of the National Academy of Sciences USA, 94, 12699±12704. Gall, J. (1977). Systemantics. New York: New York Times. Gallup, G. G., Jr. (1977). Self-recognition in primates: A comparative approach to the bi-directional properties of consciousness. American Psychologist, 32, 329± 338. Gammaitoni, L., HaÈnggi, P., Jung, P., and Marchesoni, F. (1998). Stochastic resonance. Reviews of Modern Physics, 70, 223±287. Garner, W. R. (1962). Uncertainty and Structure as Psychological Concepts. New York: Wiley. Geschieder, G. A. (1997). Psychophysics: The Fundamentals. 3d ed. Mahwah, NJ: Erlbaum. Gilden, D. L. (1997). Fluctuations in the time required for elementary decisions. Psychological Science, 8, 296±301. Gilden, D. L., Thornton, T., and Mallon, M. W. (1995). 1=f noise in human cognition. Science, 267, 1837±1839. Gleick, J. (1987). Chaos: Making a New Science. New York: Viking. Gluckman, B. J., Neto¨, T. I., Neel E. J., Ditto, W. L., Spano, M. L., and Schi¨, S. J. (1996). Stochastic resonance in a neuronal network from mammalian brain. Physical Review Letters, 77, 4098±4101. È ber formal unentscheidbare SaÈtze der Principia Mathematica GoÈdel, K. (1931). U und Verwandter Systeme I. Monatskefte fuÈr Mathematik und Physik, 38, 173± 198. Gold, T. (1965) The arrow of time. In S. T. Butler and H. Messel, eds., Time: Selected Lectures, 143±165. Oxford: Pergamon Press. Goldberg, S. (1986). Introduction to Di¨erence Equations. New York: Dover. Golden, R. M. (1996). Mathematical Methods for Neural Network Analysis and Design. Cambridge, MA: MIT Press. Goodfellow, L. D. (1938). A psychological interpretation of the results of the Zenith Radio Experiments in telepathy. Journal of Experimental Psychology, 23, 601±632. Goss-Custard, J. D. (1977). The energetics of prey selection by redshank, Tringa totanus (L.), in relation to prey density. Journal of Animal Ecology, 46, 1±19. Gottman, J. M. (1999). The Marriage Clinic: A Scienti®cally Based Marital Therapy. New York: Norton. Gottman, J. M., and Roy, A. K. (1990). Sequential Analysis: A Guide for Behavioral Researchers. New York: Cambridge University Press. Graf, P. (1994). Explicit and implicit memory: A decade of research. In C. UmiltaÁ and M. Moscovitch, eds., Attention and Performance 15: Conscious and Nonconscious Information Processing, 682±692. Cambridge, MA: MIT Press.

References

331

Granger, C. W. J. (1980). Long memory relationships and the aggregation of dynamic models. Journal of Econometrics, 14, 227±238. Granger, C. W. J., and Ding, Z. (1996). Varieties of long memory models. Journal of Econometrics, 73, 61±77. Grassberger, P., and Procaccia, I. (1983a). Measuring the strangeness of strange attractors. Physica D, 9, 189±208. Grassberger, P., and Procaccia, I. (1983b). Characterization of strange attractors. Physical Review Letters, 50, 346. Gratton, G., and Fabiani, M. (1998). Dynamic brain imaging: Event-related optical signal (EROS) measures of the time course and localization of cognitiverelated activity. Psychonomic Bulletin and Review, 5, 535±563. Gray, C. M., and W. Singer. (1989). Stimulus-speci®c neuronal oscillations in orientation columns of cat visual cortex. Proceedings of the National Academy of Sciences USA, 91, 6339±6343. Green, C. D. (1998). Are connectionist models theories of cognition? Psycholoquy, 9(4). Green, D. M., and Swets, J. A. (1966). Signal Detection Theory and Psychophysics. Reprint, New York: Krieger, 1974. Greenwood, P. E. (1997). Lecture notes for Mathematics 608. Vancouver, Canada: University of British Columbia. Greenwood, P. E., and Ward, L. M. (2001). 1=f noise and autoregressive processes. Unpublished manuscript, University of British Columbia. Greenwood, P. E., Ward, L. M., Russell, D. F., Neiman, A., and Moss, F. (2000). Stochastic resonance enhances the electrosensory information available to paddle®sh for prey capture. Physical Review Letters, 84, 4773±4776. Greenwood, P. E., Ward, L. M., and Wefelmeyer, W. (1999). Statistical analysis of stochastic resonance in a simple setting. Physical Review E, 60, 4687±4695. Gregg, L. W., and Simon, H. A. (1967). An information-processing explanation of one-trial and incremental learning. Journal of Verbal Learning and Verbal Behavior, 6, 780±787. Gregson, R. A. M. (1983). Time Series in Psychology. Hillsdale, NJ: Erlbaum. Gregson, R. A. M. (1988). Nonlinear Psychophysical Dynamics. Hillsdale, NJ: Erlbaum. Gregson, R. A. M. (1992). N-Dimensional Nonlinear Psychophysics: Theory and Case Studies. Hillsdale, NJ: Erlbaum. Grimmett, G., and Stirzaker, D. (1992). Probability and Random Processes. 2d ed. New York: Oxford University Press. Grossberg, S. (1988). Nonlinear neural networks: Principles, mechanisms, and architectures. Neural Networks, 1, 17±61. Haken, H. (1983). Synergetics. New York: Springer. Haken, H., Kelso, J. A. S., and Bunz, H. (1985). A theoretical model of phase transitions in human hand movements. Biological Cybernetics, 51, 347±356.

332

References

Hall, E. T. (1983). The Dance of Life. Garden City, NY: Doubleday. Hamero¨, S. R. (1998). ``Funda-mentality'': Is the conscious mind subtly linked to a basic level of the universe? Trends in Cognitive Sciences, 2, 119±124. Hansel, C. E. M. (1966). ESP: A Scienti®c Evaluation. New York: Scribner. Hausdor¨, J. M., and Peng, C. K. (1996). Multiscaled randomness: A possible source of 1=f noise in biology. Physical Review E, 54, 2154±2157. Hebb, D. O. (1949). The Organization of Behavior. New York: Wiley. Hecht, S., Schlaer, S., and Pirenne, M. H. (1942). Energy, quanta, and vision. Journal of General Physiology, 25, 819±840. He¨ernan, M. (1996). Comparative e¨ects of microcurrent stimulation on EEG. Integrative Physiological and Behavioral Science, 31, 202±209. Heisenberg, W. (1927). The physical content of quantum kinematics and mechanics. Zeitschrift fuÈr Physik, 43, 172. Herrnstein, R. J. (1961). Relative and absolute strength of response as a function of frequency of reinforcement. Journal of the Experimental Analysis of Behavior, 4, 267±272. Hillyard, S. A., and Picton, T. W. (1987). Electrophysiology of cognition. In F. Plum, ed., Handbook of Physiology: The Nervous System. Vol. 5, Higher Functions of the Nervous System, 519±584. Bethesda, MD: American Physiological Society. Hodgkin, A. L., and Huxley, A. F. (1952). A quantitative description of membrane current and its application to conduction and excitation in nerve. Journal of Physiology, 117, 500±544. Hofstadter, D. R. (1979). GoÈdel, Escher, Bach: An Eternal Golden Braid. New York: Basic Books. Honerkamp, J. (1998). Statistical Physics: An Advanced Approach with Applications. New York: Springer. Houghton, G., and Tipper, S. P. (1994). A model of inhibitory mechanisms in selective attention. In D. Dagenbach and T. H. Carr, eds., Inhibitory Processes in Attention, Memory, and Language, 53±112. San Diego, CA: Academic Press. Houston, C. E. (1999). Thought suppression of episodic memories and susceptibility to erroneous suggestion regarding sequence. Dissertation Abstracts International: Section B: The Sciences and Engineering, 60(1-B), 0411. Hsieh, D. (1991). Chaos and nonlinear dynamics: Application to ®nancial markets. Journal of Finance, 46, 1839±1877. Hudson, D. L., and Cohen, M. E. (2000). Neural Networks and Arti®cial Intelligence for Biomedical Engineering. New York: Institute of Electrical and Electronics Engineers. Iberall, A. S. (1972). Toward a General Science of Viable Systems. New York : McGraw-Hill. Ja¨e, J., and Feldstein, S. (1970). Rhythms of Dialogue. New York: Academic Press.

References

333

James, W. (1890). The Principles of Psychology in Two Volumes. Vol. 1. New York: Holt. JasÂkowski, P., and Verleger, R. (1999). Amplitudes and latencies of single-trial ERPs estimated by a maximum likelihood method. IEEE Transactions on Biomedical Engineering, 46, 987±993. JasÂkowski, P., and Verleger, R. (2000). An evaluation of methods for single-trial estimation of P3 latency. Psychophysiology, 37, 153±162. Jaynes, J. (1976). The Origin of Consciousness in the Breakdown of the Bicameral Mind. Boston: Houghton Mi¿in. Jesteadt, W., Luce, R. D., and Green, D. M. (1977). Sequential e¨ects in judgments of loudness. Journal of Experimental Psychology: Human Perception and Performance, 3, 92±104. John, E. R. (1977). Neurometrics. Science, 196, 1393±1410. Jourdain, R. (1997). Music, the Brain, and Ecstasy. New York: Avon. Just, M. A., and Carpenter, P. A. (1976). Eye ®xations and cognitive processes. Cognitive Psychology, 8, 441±480. Just, M. A., and Carpenter, P. A. (1987). The Psychology of Reading and Language Comprehension. Boston: Allyn and Bacon. Kac, M. (1983). What is random? American Scientist, 71, 405±406. Kac, M. (1984). More on randomness. American Scientist, 72, 282±283. Kahneman, D. (1968). Method, ®ndings, and theory in studies of visual masking. Psychological Bulletin, 70, 404±425. Kahneman, D., and Tversky, A. (1982). The simulation heuristic. In Kahneman, P. Slovic, and Tversky, eds., Judgement under Uncertainty: Heuristics and Biases, 201±208. Cambridge: Cambridge University Press. Kalmus, M., and Bachmann, T. (1980). Perceptual microgenesis of complex visual pattern: Comparison of methods and possible implications for future studies. Acta et Commentationes Universitatis Tartuensis, 529, 134±159. Kandel, E. R., Schwartz, J. H., and Jessell, T. M., eds. (2000). Principles of Neural Science. 4th ed. New York: McGraw-Hill. Kareev, Y. (1992). Not that bad after all: Generation of random sequences. Journal of Experimental Psychology: Human Perception and Performance, 18, 1189±1194. Kau¨man, H., and Noyes, H. P. (1996). Discrete physics and the derivation of electromagnetism from the formalism of quantum mechanics. Proceedings of the Royal Society of London A, 452, 81±95. Kau¨man, S. A. (1992). Applied molecular evolution. Journal of Theoretical Biology, 157, 1. Kau¨man, S. A. (1993). The Origins of Order. New York: Oxford University Press. Kaulakys, B. (1999). Autoregressive model of 1=f noise. Physics Letters A, 257, 37±42.

334

References

Kaulakys, B., and MesÏauskas, T. (1998). Modeling 1=f noise. Physical Review E, 58, 7013±7019. Kellert, S. H. (1993). In the Wake of Chaos. Chicago: University of Chicago Press. Kelso, J. A. S. (1995). Dynamic Patterns: The Self-Organization of Brain and Behaviour. Cambridge, MA: MIT Press. Kennel, M. B., and Isabelle, S. (1992). Method to distinguish possible chaos from colored noise and to determine embedding parameters. Physical Review A, 46, 3111±3118. Killeen, P. R. (1989). Behavior as a trajectory through a ®eld of attractors. In J. R. Brink and C. R. Haden, eds., The Computer and the Brian: Perspectives on Human and Arti®cial Intelligence, 53±82. New York: Elsevier Science Publishers. Kilmister, C. W. (1992). Space, time, discreteness. Philosophica, 50, 55±71. Kinsbourne, M. (1988). Integrated ®eld theory of consciousness. In A. J. Marcel and E. Bisiach, eds., Consciousness in Contemporary Science, 239±256. New York: Oxford University Press. Klein, R. M. (2000). Inhibition of return. Trends in Cognitive Sciences, 4, 138± 147. Kleinmuntz, B. (1968). The processing of clinical information by man and machine. In B. Kleinmuntz, ed., Formal Representation of Human Judgment. New York: Wiley. Klir, G. J. (1972a). Trends in General Systems Theory. New York: Wiley. Klir, G. J. (1972b). Introduction to the Methodology of Switching Circuits. New York: Van Nostrand. Koch, C. (1997). Computation and the single neuron. Nature, 385, 207±210. Koch, C., and Segev, I. (1998). Methods in Neuronal Modeling: From Ions to Networks. 2d ed. Cambridge, MA: MIT Press. KoÈnig, P., Engel, A. K., and Singer, W. (1996). Integrator or coincidence detector? The role of the cortical neuron revisited. Trends in Neuroscience, 19, 130±137. Kopell, N. (1988). Toward a theory of modelling central pattern generators. In A. H. Cohen, S. Rossignol, and S. Grillner, eds., Neural Control of Rhythmic Movements in Vertebrates, 369±413. New York: John Wiley & Sons. Kosslyn, S. M., and Koenig, O. (1992). Wet Mind: The New Cognitive Neuroscience. New York: Free Press. Kramers, H. A. (1940). Brownian motion in a ®eld of force and the force and the di¨usion model of chemical reactions. Physica, 7, 284±304. Krebs, J. R., Kacelnik, A., and Taylor, P. (1978). Optimal sampling by birds: An experiment with great tits (Parus major). Nature, 275, 27±31. Kubvoy, M., Rapoport, A., and Tversky, A. (1970). Deterministic vs. Probabilistic strategies in detection. Perception and Psychophysics, 9, 427±429. Kurthen, M., Grunwald, T., and Elger, C. E. (1998). Will there be a neuroscienti®c theory of consciousness? Trends in Cognitive Sciences, 2, 229±234.

References

335

Lamme, V. A. F., and Roelfsema, P. R. (2000). The distinct modes of vision o¨ered by feedforward and recurrent processing. Trends in Neurosciences, 23, 571±579. Large, E. W., and Jones, M. R. (1999). The dynamics of attending: How people track time varying events. Psychological Review, 106, 119±159. Lashley, K. S. (1950). In search of the engram. Symposia of the Society for Experimental Biology, 4, 454±482. Lashley, K. S. (1951). The problem of serial order in behavior. In L. A. Je¨ress, ed., Cerebral Mechanisms in Behavior, 112±136. New York: Wiley. Lauk, M., Chow, C. C., Pavlik, A. E., and Collins, J. J. (1998). Human balance out of equilibrium: Nonequilibrium statistical mechanics in posture control. Physical Review Letters, 80, 413±416. Lauritzen, S. L. (1996). Graphical Models. Oxford Statistical Science Series, vol. 17. New York: Oxford University Press. Lea, S. E. G. (1984). Instinct, Environment and Behaviour. London: Methuen. Lee, W. (1971). Decision Theory and Human Behavior. New York: Wiley. Leibniz, G. F. (1714/1956). Monadologie. Paris: Delagrave. Leipus, R., and Viano, M. C. (2000). Modeling long-memory time series with ®nite or in®nite variance: A general approach. Journal of Time Series Analysis, 21, 61±74. Levelt W. J. M. (1968). On Binocular Rivalry. Hague: Mouton. Levin, J. E., and Miller, J. P. (1996). Broadband neural encoding in the cricket cercal sensory system enhanced by stochastic resonance. Nature, 380, 165±168. Lewis, F. L., Jagannathan, S., and YesËildirek, A. (1999). Neural Network Control of Robot Manipulators and Nonlinear Systems. London: Taylor and Francis. Liberman, A. M., and Mattingly, I. G. (1985). The motor theory of speech perception revised. Cognition, 21, 1±36. Link, S. W. (1992). The Wave Theory of Di¨erence and Similarity. Hillsdale, NJ: Erlbaum. Link, S. W. (1994). Rediscovering the past: Gustav Fechner and signal detection theory. Psychological Science, 5, 335±340. Lisman, J. E., and Idiart, M. A. P. (1995). Storage of 7 G 2 short-term memories in oscillatory subcycles. Science, 267, 1512±1515. Lockhead, G. R. (1974). Sequential behavior and parapsychology. Unpublished manuscript, Duke University. Logothetis, N. K., and Schall, J. D. (1989). Neuronal correlates of subjective visual perception. Science, 245, 761±763. Longtin, A. (1993). Stochastic resonance in neuron models. Journal of Statistical Physics, 70, 309±327. Longtin, A. (1995a). Synchronization of the stochastic FitzHugh-Nagumo Equations to periodic forcing. Il Nuovo Cimento, 17D, 835±846.

336

References

Longtin, A. (1995b). Mechanisms of stochastic phase locking. Chaos, 5(1), 209± 215. Longtin, A. (1997). Autonomous stochastic resonance in bursting neurons. Physical Review E, 55, 1±9. Luce, R. D. (1986). Response Times: Their Role in Inferring Elementary Mental Organization. Oxford Psychology Series, vol. 8. New York: Oxford University Press. Lumer, E. D., Edelman, G. M., and Tononi, G. (1997). Neural dynamics in a model of the thalamocortical system: l Layers, loops and the emergence of fast synchronous rhythms. Cerebral Cortex, 7, 207±227. LundstroÈm, I., and McQueen, D. (1974). A proposed 1=f noise mechanism in nerve cell membranes. Journal of Theoretical Biology, 45, 405±409. Luria, S. E., Gould, S. J., and Singer, S. (1981) A View of Life. Menlo Park, CA: Benjamin-Cummings. Machlup, S. (1981). Earthquakes, thunderstorms, and other 1=f noises. In P. H. E. Meijer, R. D. Mountain, and R. J. Soulen, Jr., eds., Sixth International Conference on Noise in Physical Systems, 157±160. Washington, DC: National Bureau of Standards. Macmillan, N. A., and Creelman, C. D. (1991). Detection Theory: A User's Guide. New York: Cambridge University Press. Mandelbrot, B. B. (1982). The Fractal Geometry of Nature. New York: Freeman. Mandelbrot, B. B. (1998). Multifractals and 1=f Noise: Wild Self-A½nity in Physics. New York: Springer. Mandler, G. (1980). Recognizing: The judgment of previous occurrences. Psychological Review, 87, 252±271. Marcel, A. J. (1983). Conscious and unconscious perception: Experiments on visual masking and word recognition. Cognitive Psychology, 15, 197±237. Marks, L. E. (1974). Sensory Processes: The New Psychophysics. New York: Academic Press. Maslov, S., Tang, C., and Zhang, Y. C. (1999). 1=f noise in Bak-Tang-Wiesenfeld models on narrow stripes. Physical Review Letters, 83, 2449±2452. Mato, G. (1998). Stochastic Resonance in neural systems: E¨ect of temporal correlation of spike trains. Physical Review E, 58, 876±880. Mayer-Kress, G., ed. (1986). Dimensions and Entropies in Chaotic Systems. New York: Springer. McCleary, R., and Hay, R. A., Jr. (1980). Applied Time Series Analysis for the Social Sciences. Beverly Hills, CA: Sage. McDonald, J. J., and Ward, L. M. (1998). Nonlinear dynamics of event-related human brain activity. Unpublished manuscript, University of British Columbia. Meadows, D. H., Meadows, D. L., Randers, J., and Behrens, W. W., III. (1972). The Limits to Growth. Washington, DC: Potomac.

References

337

Menning, H., Roberts, L. E., and Pantev, C. (2000). Plastic changes in the auditory cortex induced by intensive frequency discrimination training. Neuroreport, 11, 817±822. Merriam-Webster. (1971). The New Merriam-Webster Dictionary. New York: Pocket Books. Metcalfe, J. (1986). Premonitions of insight predict impending error. Journal of Experimental Psychology: Learning Memory and Cognition, 12, 623±634. Metcalfe, J., and Wiebe, D. (1987). Intuition in insight and noninsight problem solving. Memory and Cognition, 15, 238±246. Metzger, M. A. (1994). Have subjects been shown to generate chaotic numbers? Commentary on Neuringer and Voss. Psychological Science, 5, 111±114. Miller, G. A. (1952). Finite Markov processes in psychology. Psychometriks, 17, 149±167. Miller, G. A. (1956). The magical number seven, plus or minus two: Some limits on our capacity for processing information. Psychological Review, 63, 81±97. Milotti, E. (1995). Linear processes that produce 1=f or ¯icker noise. Physical Review E, 51, 3087±3103. Minelli, A. (1971). Memory, morphogenesis and behavior. Scientia, 106, 798±806. Minsky, M. L., and Papert, S. A. (1967). Perceptrons. Cambridge, MA: MIT Press. Molnar, M., and Skinner, J. E. (1992). Low-dimensional chaos in event-related brain potentials. International Journal of Neuroscience, 66, 263±276. Morrison, F. (1991). The Art of Modeling Dynamic Systems: Forecasting for Chaos, Randomness, and Determinism. New York: Wiley-Interscience. Morse, R. P., and Evans, E. F. (1996). Enhancement of vowel coding for cochlear implants by addition of noise. Nature Medicine, 2, 928±932. Moss, F., Pierson, D., and O'Gorman, D. (1994). Stochastic resonance: Tutorial and update. International Journal of Bifurcation and Chaos, 4, 1383±1397. Moss, F., Chiou-Tan, F., and Klinke, R. (1996). Will there be noise in their ears? Nature Medicine, 2, 860±862. Moss, F., Douglass, J. K., Wilkens, L., Pierson, D., and Pantazelou, E. (1993). Stochastic resonance in an electronic FitzHugh-Nagumo model. In J. R. Buchler and H. E. Kandrup, eds., Stochastic Processes in Astrophysics, Annals of the New York Academy of Sciences, 706, 26±41. MuÈller, U. U. (2000). Nonparametric regression for threshold data. Canadian Journal of Statistics, 28, 301±310. MuÈller, U. U., and Ward, L. M. (2000). Stochastic resonance in a statistical model of a time-integrating detector. Physical Review E, 61, 4286±4294. Musha, T. (1981). 1=f ¯uctuations in biological systems. In P. H. E. Meijer, R. D. Mountain, and R. J. Soulen, Jr., eds., Sixth International Conference on Noise in Physical Systems, 143±146. Washington, DC: U.S. Department of Commerce and National Bureau of Standards.

338

References

NaÈaÈtaÈnen, R. (1992). Attention and Brain Function. Hillsdale, NJ: Erlbaum. Nagumo, J., Arimoto, S., and Yoshizawa, S. (1962). An active pulse transmission line simulating nerve axon. Proceedings of the Institute of Radio Engineers, 50, 2061±2070. Nakao, H. (1998). Asymptotic power law of moments in a random multiplicative process with weak additive noise. Physical Review E, 58, 1591±1600. Nass, G. (1970). The Molecules of Life. New York: McGraw-Hill. Neisser, U. (1981). John Dean's memory: A case study. Cognition, 9, 1±22. Neuringer, A. (1986). Can people behave ``randomly''? The role of feedback. Journal of Experimental Psychology: General, 115, 62±75. Neuringer, A., and Voss, C. (1993). Approximating chaotic behavior. Psychological Science, 4, 113±119. Newell, A. (1990). Uni®ed Theories of Cognition. Cambridge MA: Harvard University Press. Newell, A. (1992). PreÂcis of Uni®ed Theories of Cognition. Behavioural and Brain Sciences, 15, 425±492. Nordsieck, A., Lamb, W. E., and Uhlenbeck, G. E. (1940). On the theory of cosmic-ray showers: 1. The Furry model and the ¯uctuation problem. Physica, 7, 344±360. Norman, D. A. (1982). Learning and Memory. San Francisco: Freeman. Norwich, K. H. (1993). Information, Sensation, and Perception. Toronto: Academic Press. Novikov, E., Novikov, A., Shannaho¨-Khalsa, D., Schwartz, B., and Wright, J. (1997). Scale-similar activity in the brain. Physical Review E, 56, R2387± R2389. Nunez, P. L. (1995). Neocortical Dynamics and Human EEG Rhythms. New York: Oxford University Press. Nunez, P. L. (2000). Toward a quantitative description of large-scale neocortical dynamic function and EEG. Behavioral and Brain Sciences, 23, 371±437. Osborne, A. R., and Provenzale, A. (1989). Finite correlation dimension for stochastic systems with power law spectra. Physica D, 35, 357±381. Packard, N. H, Crutch®eld, J. P., Farmer, J. D., and Shaw, R. S. (1980). Geometry from a time series. Physics Review Letters, 45, 712±716. Pacut, A. (1978). Stochastic model of the latency of the conditioned escape response. Progress in Cybernetics and System Research, 3, 633±642. Pantev, C., Oostenveld, R., Engelien, A., Ross, B., Roberts, L. E., and Hoke, M. (1998). Increased auditory coritcal represenation in musicians. Nature, 392, 811± 814. Pantev, C., Wollbrink, A., Roberts, L. E., Engelein, A., and LutkenhoÈner, B. (1999). Short-term plasticity of the human auditory cortex. Brain Research, 842, 192±199.

References

339

Parks, R. W., Levine, D. S., and Long, D. L. (1998). Fundamentals of Neural Network Modeling: Neuropsychology and Cognitive Neuroscience. Cambridge, MA: MIT Press. Penrose, R. (1990). The Emperor's New Mind. Oxford: Oxford University Press. Penrose, R. (1994). Shadows of the Mind: A Search for the Missing Science of Consciousness. Oxford: Oxford University Press. Pezard, L., and Nandrino, J.-L. (1996). Depression as a dynamical disease. Biological Psychiatry, 39, 991±999. Pfeifer, R., and Sheier, C. (1999). Understanding Intelligence. Cambridge, MA: MIT Press. Piaget, J. (1954). The Construction of Reality in the Child. Cambridge, MA: MIT Press. Picton, T. W., Lins, O. G., and Scherg, M. (1995). The recording and analysis of event-related potentials. Handbook of Neuropsychology, 10, 3±73. Platt, J. (1964). Strong inference. Science, 146, 347±353. Plesser, H. E., and Tanaka, S. (1997). Stochastic resonance in a model neuron with reset. Physics Letters A, 225, 228±234. Plischke, M., and Bergersen, B. (1994). Equilibrium Statistical Physics. 2d ed. Singapore: World Scienti®c. Pons, T. P., Garraghty, P. E., Ommaya, A. K., Kaas, J. H., Taub, E., and Mishkin, M. (1991). Massive cortical reorganization after sensory dea¨erentation in adult macaques. Science, 252, 1857±1860. Port, R. F., and van Gelder, T. (1995). Mind as Motion: Explorations in the Dynamics of Cognition. Cambridge, MA: MIT Press. Posner, M. I., and Raichle, M. E. (1994). Images of Mind. New York: Scienti®c American Library. Povel, D. J. (1981). Internal representation of simple temporal patterns. Journal of Experimental Psychology: Human Perception and Performance, 7, 3±18. Povinelli, D. J. (1993). Reconstructing the evolution of the mind. American Psychologist, 48, 493±509. Press, W. H., Teykolsky, S. A., Vetterling, W. T., and Flannnery, B. P. (1992). Numerical Recipes in C: The Art of Scienti®c Computing. 2d ed. New York: Cambridge University Press. Pribram, K. H. (1986). The cognitive revolution and mind/brain issues. American Psychologist, 41, 507±520. Pribram, K. H. (1996). Chaos edge. In W. Sulis and A. Combs, eds., Nonlinear Dynamics in Human Behavior, v±x. Singapore: World Scienti®c. Prigogene, I., and Stengers, I. (1984). Order Out of Chaos. Toronto: Bantam. Pritchard, W. S., and Duke, D. W. (1992). Measuring chaos in the brain: A tutorial review of nonlinear dynamical EEG analysis. International Journal of Neuroscience, 67, 31±80.

340

References

Ramirez, R. W. (1985). The FFT: Fundamentals and Concepts. Englewood Cli¨s, NJ: Prentice-Hall. Rand, R. H., Cohen A. H., and Holmes, P. J. (1988). Systems of coupled oscillators as models for central pattern generators. In Cohen, A. H., Rossignol, S., and Grillner, S., eds., Neural Control of Rhythmic Movements in Vertebrates 333± 368. New York: John Wiley & Sons. Rapoport, A. (1972). The uses of mathematical isomorphism in general systems theory. In G. J. Klir ed., Trends in General System Theory, 42±77. New York: Wiley-Interscience. Rapp, P. E. (1993). Chaos in the neurosciences: Cautionary tales from the frontier. Biologist, 40, 89±94. Rapp, P. E., Bashore, T. R., Martinerie, J. M., Albano, A. M., and Meos, A. I. (1989). Dynamics of brain electrical activity. Brain Topography, 2, 99±118. Reichenbach, H. (1949). The Theory of Probability. (E. Hutton and M. Reichenbach, Trans.) Berkeley: University of California Press. (Originally published as Wahrscheinlichkeitslehre in 1934). Rinkus, G. J. (1996). A combinatorial neural network exhibiting episodic and semantic memory properties for spatio-temporal patterns. Dissertation Abstracts International: Section B: The Sciences and Engineering, 57(5-B), 3428. Roediger, H. L., III. (1996). Memory illusions. Journal of Memory and Language, 35, 76±100. Rumelhart, D. E., Hinton, G. E., and McClelland, J. L. (1986). A general framework for parallel distributed processing. In Rumelhart, D. E. and McClelland, J. L., eds., Parallel Distributed Processing: Explorations in the Microstructure of Cognition. Vol. 1, Foundations, 45±77. Cambridge: MIT Press. Rumelhart, D. E., and McClelland, J. L., eds. (1986). Parallel Distributed Processing: Explorations in the Microstructure of Cognition. Vol. 1, Foundations. Cambridge, MA: MIT Press. Russell, B. (1945). A History of Western Philosophy. New York: Simon and Schuster. Russell, D. F., Wilkens, L. A., and Moss, F. (1999). Use of behavioral stochastic resonance by paddle®sh for feeding. Nature, 402, 291±294. Sandusky, A. (1971). Signal recognition models compared for random and Markov presentation sequences. Perception and Psychophysics, 10, 339±347. Sattath, S., and Tversky, A. (1977). Additive similarity trees. Psychometrika, 42, 319±345. SauveÂ, K. (1999). Gamma-band synchronous oscillations: Recent evidence regarding their functional signi®cance. Consciousness and Cognition, 8, 213±224. Scheier, C., and Tschacher, W. (1996). Appropriate algorithms for nonlinear time series anaylsis in psychology. In W. Sulis, A. Combs, eds., Nonlinear Dynamics in Human Behavior, 27±43. Singapore: World Scienti®c. Scholem, G. G. (1965). On the Kabbalah and its Symbolism. New York: Schocken.

References

341

È ber spontane Stromschwankungen in verschiedenen ElekSchottky, W. (1918). U trizitaÈtsleitern. Annalen der Physik, 57, 541±567. SchroÈdinger, E. (1944). What is Life? and Mind and Matter. Reprint, Cambridge: Cambridge University Press, 1967. Schroeder, M. (1991). Fractals, Chaos, Power Laws: Minutes from an In®nite Paradise. New York: Freeman. Seamans, J. K., Gorelova, N., and Yang, C. R. (1997). Contributions of voltagegated Ca 2‡ channels in the proximal versus distal dendrites to synaptic integration in prefrontal cortical neurons. Journal of Neuroscience, 17, 5936±5948. Shallice, T. (1972). Dual functions of consciousness. Psychological Review, 79, 383±393. Shannon, C. E., and Weaver, W. (1949). The Mathematical Theory of Communication. Urbana: University of Illinois Press. Sheinberg, D. L., and Logothetis, N. K. (1997). The role of temporal cortical areas in perceptual organization. Proceedings of the National Academy of Sciences USA, 94, 3408±3413. Sherrington, C. S. (1940). Man and His Nature. Cambridge: Cambridge University Press. Siegel, S. (1956). Nonparametric Methods for Behavioral Sciences. New York: McGraw-Hill. Siegelmann, H. T. (1995). Computation beyond the Turing limit. Science, 268, 545±548. Simon, H. A. (1969). The Sciences of the Arti®cial. Cambridge, MA: MIT Press. Simon, H. A. (1974). How big is a chunk? Science, 183, 482±488. Simon, H. A., and Newell, A. (1971). Human Problem Solving. Englewood Cli¨s, NJ: Prentice Hall. Simonotto, E., Riani, M., Seife, C., Roberts, M., Twitty, J., and Moss, F. (1997). Visual perception of stochastic resonance. Physical Review Letters, 78, 1186±1189. Singer, W. (1993). Synchronization of cortical activity and its putative role in information processing and learning. Annual Review of Physiology, 55, 349±374. Singer, W. (1994). The organization of sensory motor representations in the neocortex: A hypothesis based on temporal coding. In C. UmiltaÁ, and M. Moscovitch, eds., Attention and Performance 15: Conscious and Nonconscious Information Processing, 77±107. Cambridge, MA: MIT Press. Singer, W. (1999). Neuronal synchrony: A versatile code for the de®nition of relations? Neuron, 24, 49±65. Skarda, C. A., and Freeman, W. J. (1987). How brains make chaos in order to make sense of the world. Behavioral and Brain Sciences, 10, 161±195. Skinner, B. F. (1974). About Behaviorism. New York: Knopf. Sokolov, E. N. (1975). The neuronal mechanisms of the orienting re¯ex. In Sokolov and O. S. Vinogradova, eds., Neuronal Mechanisms of the Orienting Re¯ex, 217±238. New York: Wiley.

342

References

Squire, L. R. (1987). Memory and Brain. New York: Oxford University Press. Stanley, S. M. (1981). The New Evolutionary Timetable: Fossils, Genes, and the Origin of Species. New York: Basic Books. Stemmler, M. (1996). A single spike su½ces: The simplest form of stochastic resonance in model neurons. Network: Computation in Neural Systems, 7, 687±716. Stephens, D. W., and Krebs, J. R. (1986). Foraging Theory. Princeton: Princeton University Press. Sternberg, R. J. (1986). A triangular theory of love. Psychological Review, 93, 119±135. Sternberg, S. (1966). High-speed scanning in human memory. Science, 153, 652± 654. Stevens, S. S. (1975). Psychophysics: Introduction to its Perceptual, Neural, and Social Prospects. New York: Wiley-Interscience. Stewart, I. (1989). Does God Play Dice? The Mathematics of Chaos. Cambridge, MA: Basil Blackwell. Sugihara, G., and May, R. M. (1990). Nonlinear forecasting as a way of distinguishing chaos from measurement error in time series. Nature, 344, 734±741. Swindale, N. V. (1990). Is the cerebral cortex modular? Trends in Neuroscience, 13, 487±492. Swindale, N. V. (1998). Cortical organization: Modules, polymaps and mosaics. Current Biology, 8, R270±R273. Takakura, K., Sano, K., Kosugi, Y., and Ikebe, J. (1979). Pain control by electrical nerve stimulating using irregular pulse of 1=f ¯uctuation. Applied Neurophysiology, 42, 314±315. Takens, F. (1981). Detecting strange attractors in ¯uid turbulence. In D. A. Rand and L.-S. Young, eds., Dynamic Systems and Turbulence, 366±381. New York: Springer. Tegmark, M. (2000). Importance of quantum decoherence in brain processes. Physical Review E, 61, 4194±4206. Theiler, J. (1991). Some comments on the correlation dimension of 1=f z noise. Physics Letters A, 155, 480±493. Theiler, J., and Rapp, P. E. (1996). Re-examination of the evidence for lowdimensional, nonlinear structure in the human electroencephalogram. Electroencephalography and Clinical Neurophysiology, 98, 213±222. Theiler, J., Eubank, S., Longtin, A., Galdrikian, B., and Farmer, J. D. (1992). Testing for nonlinearity in time series: The method of surrogate data. Physica D, 58, 77±94. Thelen, E., SchoÈner, G., Schier, C., and Smith, L. B. (2001). The dynamics of embodiment: A ®eld theory of infant perseverative reaching. Behavioral and Brian Sciences, 24. Thompson, J. M. T., and Stewart, H. B. (1986). Nonlinear Dynamics and Chaos: Geometrical Methods for Engineers and Scientists. Chichester, U.K.: Wiley.

References

343

Thomson, A. M., West, D. C., Hahn, J., and Deuchars, J. (1996). Single axon IPSPs elicited in pyramidal cells by three classes of interneurons in slices of rat neocortex. Journal of Physiology (London), 496, 81±102. t'Hooft, G. (1989). On the quantization of space and time. In S. Deser, ed., Themes in Contemporary Physics II: Essays in Honor of Julian Schwinger's Seventieth Birthday, 77±89. Singapore: World Scienti®c. Thorne, K. (1994). Black Holes and Time Warps: Einstein's Outrageous Legacy. New York: Norton. Thurner, S., Feurstein, M. C., and Teich, M. C. (1997). Conservation laws in coupled multiplicative random arrays lead to 1=f noise. Unpublished manuscript. Thurstone, L. L. (1927). A law of comparative judgment. Psychological Review, 34, 273±286. Tipler, F. J. (1994). The Physics of Immortality. New York: Doubleday. Tononi, G., and Edelman, G. M. (1998). Consciousness and complexity. Science, 282, 1846±1851. Tononi, G., Srinivasan, R., Russell, D. P., and Edelman, G. M. (1998). Investigating neural correlates of conscious perception by frequency-tagged neuromagnetic responses. Proceedings of the National Academy of Sciences USA, 95, 3198±3203. Tougaard, J. (2000). Stochastic resonance and signal detection in an energy detector-implications for biological receptor systems. Biological Cybernetics, 83, 471±480. Townsend, J. T. (1992). Chaos theory: A brief tutorial. In A. F. Healy, S. M. Kosslyn, and R. M. Shi¨rin, eds., Essays in Honor of William K. Estes, Vol. 2: From Learning Processes to Cognitive Processes, 65±96. Hillsdale, NJ: Lawrence Erlbaum Associates. Townsend, J. T., and Ashby, F. G. (1983). Stochastic Modeling of Elementary Psychological Processes. New York: Cambridge University Press. Trabasso, T., and Bower, G. (1968). Attention in Learning. New York: Wiley. Treisman, M., and Faulkner, A. (1987). Generation of random sequences by human subjects: Cognitive operations or psychophysical process? Journal of Experimental Psychology: General, 116, 337±355. Triesman, M., and Williams, T. C. (1984). A theory of criterion setting with an application to sequential dependencies. Psychological Review, 91, 68±111. Tulving, E. (1983). Elements of Episodic Memory. London: Clarendon Press. Tune, G. S. (1964). A brief survey of variables that in¯uence random generation. Perceptual and Motor Skills, 18, 705±710. Turing, A. M. (1936). On computable numbers with an application to the Entscheidungs problem. Proceedings of the London Math Society, 42, 230±265 and 43, 544±546. Tversky, A., and Kahneman, D. (1981). The framing of decisions and the psychology of choice. Science, 211, 453±458.

344

References

Usher, M., Stemmler, M., and Olami, Z. (1995). Dynamic pattern formation leads to 1=f noise in neural populations. Physical Review Letters, 74, 326±329. Uttal, W. R. (1969). Masking of alphabetic character recognition by dynamic visual noise (DVN). Perception and Psychophysics, 6, 121±127. Van Bendegem, J. P., ed. (1992). Modern Perspectives on the Philosophy of Space and Time. Philosophica, 50(2). Special issue. van der Pol, B. (1926). On ``relaxation-oscillations.'' Philosophical Magazine and Journal of Science, 2, 978±992. van der Pol, B., and van der Mark, J. (1927). Frequency demultiplication. Nature, 120, 363±364. van der Pol, B., and van der Mark, J. (1928). The heartbeat considered as a relaxation oscillation, and an electrical model of the heart. Philosophical Magazine and Journal of Science, 2, 763±775. Van Essen, D. C., and Deyoe, E. A. (1995). Concurrent processing in the primate visual cortex. In M. S. Gazzaniga, ed., The Cognitive Neurosciences, 383±400. Cambridge, MA: MIT Press. van Gelder, T. (1998). The dynamical hypothesis in cognitive science. Behavioral and Brain Sciences, 21, 615±628. van Gelder, T., and Port, R. F. (1995). It's about time: An overview of the dynamical approach to cognition. In Port and van Gelder, eds., Mind as Motion: Explorations in the Dynamics of Cognition, 1±43. Cambridge, MA: MIT Press. van Kampen, N. G. (1976). The expansion of the master equation. Advances in Chemical Physics, 34, 245±309. van Kampen, N. G. (1985). Stochastic Processes in Physics and Chemistry. Amsterdam: North-Holland. van Kampen, N. G. (1987). Some theoretical aspects of noise. In C. M. Van Vliet, ed., Ninth International Conference on Noise in Physical Systems, 3±10. Singapore: World Scienti®c. Van Vliet, K. M. (1981). Classi®cation of noise phenomena. In P. H. E. Meijer, R. D. Mountain, and R. J. Soulen, Jr., eds., Sixth International Conference on Noise in Physical Systems, 3±11. Washington, DC: U.S. Department of Commerce and National Bureau of Standards. von Bertelan¨y, L. (1968). General Systems Theory. New York: George Braziller. von Bertelan¨y, L. (1972). The history and status of general systems theory. In G. J. Klir, ed., Trends in General Systems Theory, 21±41. New York: WileyInterscience. Voss, R. F., and Clarke, J. (1975). ``1=f noise'' in music and speech. Nature, 258, 317±318 Wagenaar, W. A. (1971). Serial nonrandomness as a function of duration and monotony of a randomization task. Acta Psychologica, 35, 70±87.

References

345

Wagenaar, W. A. (1972). Generation of random sequences by human subjects: A critical survey of literature. Psychological Bulletin, 77, 65±72. Waldrop, M. M. (1992). Complexity: The Emerging Science at the Edge of Order and Chaos. New York: Simon and Schuster. Ward, L. M. (1973). Use of Markov-encoded sequential information in numerical signal detection. Perception and Psychophysics, 14, 337±342. Ward, L. M. (1979). Stimulus information and sequential dependencies in magnitude estimation and cross-modality matching. Journal of Experimental Psychology: Human Perception and Performance, 5, 444±459. Ward, L. M. (1987). Remembrance of sounds past: Memory and psychophysical scaling. Journal of Experimental Psychology: Human Perception and Performance, 13, 216±227. Ward, L. M. (1990). Critical bands and mixed-frequency scaling: Sequential dependencies, equal-loudness contours, and power function exponents. Perception and Psychophysics, 47, 551±562. Ward, L. M. (1991). Informational and neural adaptation curves are asynchronous. Perception and Psychophysics, 50, 117±128. Ward, L. M. (1992). Mind in psychophysics. In D. Algom, ed., Psychophysical Approaches to Cognition, 187±249. Amsterdam: North-Holland. Ward, L. M. (1996). Hypothesis testing, nonlinear forecasting, and the search for chaos in psychophysics. In W. Sulis and A. Combs, eds., Nonlinear Dynamics in Human Behavior, 77±89. Singapore: World Scienti®c. Ward, L. M. (1999). The psychophysics of stochastic resonance. In P. Killeen, ed., Fechner Day '99, 389±394. Tempe, AZ: International Society for Psychophysics. Ward, L. M. (2000). Signal detection theory and stochastic resonance. In C. Bonnett, ed., Fechner Day 2000, 357±362. Strasbourg: International Society for Psychophysics. Ward, L. M. (2001). Human neural plasticity. Trends in Cognitive Sciences, 5, 325±327. Ward, L. M., Livingston, J. W., and Li, J. (1988). On probabilistic categorization: The Markovian observer. Perception and Psychophysics, 43, 125±136. Ward, L. M., and Lockhead, G. R. (1971). Response system processes in absolute judgement. Perception and Psychophysics, 9, 73±78. Ward, L. M., Moss, F., Desai, S., and Rootman, D. (2001). Stochastic resonance in detection of auditory beats by humans. Unpublished manuscript, Univ. of British Columbia. Ward, L. M., Moss, F., Desai, S., and Tata, M. (2001). Stochastic resonance in detection of visual grating by humans. Unpublished manuscript, Univ. of British Columbia. Ward, L. M., Neiman, A., and Moss, F. (2001). Stochastic resonance in psychophysics and in animal behavior. Unpublished manuscript, Univ. of British Columbia.

346

References

Ward, L. M., and Richard, C. M. (2001). 1=f z noise and decision complexity. Unpublished manuscript, Univ. of British Columbia. Ward, L. M., and West, R. L. (1994). On chaotic behavior. Psychological Science, 5, 232±236. Ward, L. M., and West, R. L. (1998). Modeling human chaotic behavior: Nonlinear forecasting analysis of logistic iteration. Nonlinear Dynamics, Psychology and Life Sciences, 3, 261±282. Weaver, W. (1963). Lady Luck: The Theory of Probability. Garden City, NY: Doubleday. Wegner, D. M., Quillian, F., and Houston, C. E. (1996). Memories out of order: Thought suppression and the disturbance of sequence memory. Journal of Personality and Social Psychology, 71, 680±691. Weinberg, G. M. (1975). An Introduction to General Systems Thinking. New York: Wiley. Weissman, M. B. (1981). Survey of recent 1=f theories. In P. H. E. Meijer, R. D. Mountain, and R. J. Soulen, Jr., eds., Sixth International Conference on Noise in Physical Systems, 3±11. Washington, DC: U.S. Department of Commerce and National Bureau of Standards. Wheeler, J. A. (1990). Information, physics, quantum: The search for links. In W. Zurek, ed., Complexity, Entropy, and the Physics of Information. Redwood City, CA: Addison-Wesley. Whishaw, I. Q., and Vanderwolf, C. H. (1973). Hippocampal EEG and behavior: Changes in amplitude and frequency of RSA (theta rhythm) associated with spontaneous and learned movement patterns in rats and cats. Behavioral Biology, 8, 461±484. White, J. A. Rubenstein, J. T., and Kay, A. R. (2000). Channel noise in neurons. Trends in Neuroscience, 23, 131±137. Wiesenfeld, K., and Moss, F. (1995). Stochastic resonance and the bene®ts of noise: from ice ages to cray®sh and SQUIDs. Nature, 373, 33±36. Wise, M. N. (1979). The mutual embrace of electricity and magnetism. Science, 203, 1310±1318. Wittgenstein, L. (1961). Tractatus Logico-Philosophicus. Trans. D. F. Pears and B. F. McGuiness. London: Routledge and Kegan Paul. Woltz, D. J., Bell, B. G., Kyllonen, P. C., and Gardner, M. K. (1996). Memory for order of operations in the acquisition and transfer of sequential cognitive skills. Journal of Experimental Psychology: Learning, Memory, and Cognition, 22, 438±457. Woodworth, R. S. (1938). Experimental Psychology. New York: Holt, Rinehart and Winston. Wright, R. D., and Liley, D. T. J. (1996). Dynamics of the brain at global and microscopic scales: Neural networks and the EEG. Behavioral and Brain Sciences, 19, 285±320.

References

347

Wright, R. D., and Ward, L. M. (1998). Control of visual attention. In Wright, ed., Visual Attention, 132±186. New York: Oxford University Press. Yamada, W., Koch, C., and Adams, P. R. (1998). Multiple channels and calcium dynamics. In Koch and I. Segev, eds., Methods in Neuronal Modeling: From Ions to Networks. 2d ed., 137±170. Cambridge, MA: MIT Press. Yarbus, A. L. (1967). Eye Movements and Vision. New York: Plenum Press. Zeki, S. (1993). A Vision of the Brain. London: Blackwell. Zeng, F. -G., Fu, Q. -J., and Morse, R. (2000). Human hearing enhanced by noise. Brain Research, 869, 251±255.

This page intentionally left blank

Index

Action potential, 145±147, 150, 184±186, 188, 190, 193±195, 197±199, 280±281, 296, 300±301 Adaptation, 37±38, 280, 291 Aliasing, 32 ARIMA, 89±92, 94, 96± 98, 102, 141, 169, 229, 231±232, 234±235, 258 Arti®cial intelligence, 308, 311 Attention, 5, 32, 41, 46, 62, 67, 74, 84, 135, 163, 291, 303, 316 Attractor, 205, 207±210, 213±217, 220±221, 223, 227, 230, 239, 242, 244, 266, 271, 279, 291 Autocorrelation, 91±94, 258 Autoregression, 90±92, 94, 96±97, 102±104, 106, 140, 169±171. See also ARIMA Baggott, J. E., 55, 112, 114, 116 Bak, P., 157±159 Bell, D. A., 122, 130, 147, 155, 165, 167± 170 Binocular rivalry, 317±318 Biology, 4, 6, 55, 68, 75, 286±287 Bohr, N., 114 Bremermann limit, 46 Bringhurst, R., 3 Brock, W., 230±231 Brownian motion, 91, 100, 117±118, 125± 126, 130 Budescu, D. V., 256±259 Bunz, H., 21, 282 Calculus, 26, 33, 51, 65, 99 Casdagli, M., 219, 221, 224, 226, 231±232, 242, 269 Chalmers, D. J., 316 Chance, 13, 29, 167, 190, 245±246, 255, 261±262, 289 Chaos, 29, 59, 83, 87, 96, 103±105, 112, 125±126, 138±139, 173, 201±205, 207± 210, 213±215, 219, 221, 224, 226±227, 229±232, 234, 236±237, 239±242, 244±

245, 263±266, 269, 271±273, 276, 278± 280, 282, 291, 297±298, 305, 308, 313±314 dynamics, 237 edge of, 138, 155, 213, 237, 242, 291 Chaos theory, 112, 139, 173, 201, 229, 240, 264, 273, 280 dense periodic points, 210 map, 75, 83, 168, 202±205, 210, 214, 224, 226, 238±239, 244, 266, 269, 299, 313 sensitive dependence, 59, 189, 202, 204± 205, 210, 216, 221, 224, 230, 239, 244, 247±248, 250±251, 265 strange attractor, 208, 210, 217, 220, 224, 230, 239, 240, 242, 264, 266, 269 topological transitivity, 204, 210 Chemistry, 4, 55 Clancey, W. J., 12±13, 42 Clark, C. W., 292, 294 Clarke, J., 135±136 Clayton, K., 138±139, 141, 144, 213, 235 Clock, 18, 26±27, 29, 30, 35, 137±138, 188, 207, 273, 321 Cognitive neuroscience, 74, 146 Colored noise. See Noise, colored Complementarity, 8, 59±60, 62, 69, 87, 113± 114, 324 Complexity, 20, 45±48, 82, 173, 229, 241, 243, 283, 315, 320 Computation, 46, 163, 205, 305, 313±314, 324 Computer, 29, 42, 46, 58, 66±67, 72±73, 81, 83, 94, 110, 124, 126, 143, 157, 159, 186, 189, 190, 215±216, 224, 227, 257± 259, 264±266, 269, 271, 305±306, 308, 313± 314, 317, 320 digital, 46, 305, 308, 312±314 quantum, 314, 324 Computer simulation, 67, 83, 110, 143, 159, 186, 269 Consciousness, 10±11, 28, 30, 43, 50, 63, 75, 132, 229, 282±283, 295, 301, 314±318, 320±321, 323±324

350 Consciousness (cont.) Cartesian theater, 316±317 dynamic core, 315, 320±321, 323±324 hard problem, 316, 323 neural correlate of, 317 self, 323 Contingency, 13, 20, 247±248, 250±253 Correlation dimension. See Dimensionality, correlation Correlation integral, 230, 240±241, 243 Critical point, 104±105 Cross, D. V., 68, 97±98 Cycle, 21, 29, 32, 35±36, 40, 73±75, 160, 181, 203, 208, 219, 266, 294, 303 Darwin, C., 55, 285±287 DeCarlo, L. T., 68, 97±98 Delay map, 220 Delay vector, 220±221, 223±224, 230, 240, 243±244 Dennett, D., 75, 315±316, 320±321, 323 Derivative, 33, 56, 64, 274, 276±277, 306 Determinism, 83, 264±265 Deterministic versus stochastic plot, 224, 226, 232, 234, 271±272 Di¨erence equation, 66±67, 82±83, 101± 102, 118, 205, 291, 305, 308 logistic, 7, 66±67, 83, 104, 138, 209±210, 214±216, 235±236, 264±266, 269, 271, 288, 313 Di¨erential equation, 64±67, 83, 85, 108, 157, 194, 208, 214, 220, 273±274, 276, 280, 292, 302, 305±307 linear, 292 nonlinear, 51, 83, 202 ordinary, 305±307 partial, 56±57 Dimension. See Dimensionality Dimensional complexity, 241±242, 244 Dimensionality, 213±214, 220, 227, 230, 235, 240±244 box-counting, 214, 227 correlation, 240±241, 243±244 embedding, 220, 223, 240, 243±244 point D2, 243±244 Directed graph. See Graph theory, directed graph Double-well potential, 22±24, 176±177, 208 Douglass, J. K., 184±186 Dukas, R., 285±286, 291±292, 294 Dunne, J. W., 9±11 Ecology of cognition, 291 Economics, 7, 68 Edelman, G., 283, 315, 320±321 Edge of chaos. See Chaos, edge of

Index Electricity, 57±58 Electroencephalogram, 150, 153, 208, 226, 237±238, 240±242, 295±299, 301, 303, 318, 324 Emergence, 49±50, 288, 320 Ergodicity, 105±106 Error, 5±6, 8, 30, 32, 47, 59, 63, 68, 82±84, 97, 115±116, 139, 181±182, 205, 215, 223±224, 226, 231±232, 307, 311 Euler method, 306±307 Event-related potential, 150±151, 153, 242± 244, 298 Evolution, 4, 20, 26, 28±29, 43, 47, 55, 110, 112, 116, 118, 190, 214, 216, 251, 285± 289, 291, 294, 306, 323 ®tness, 252, 287±288, 291 genes, 252, 287±289, 291 macro, 288 natural selection, 55, 285±287, 291 punctuated equilibrium, 288 Exceedance, 179, 182 Explicit memory. See Memory, explicit Exponential distribution. See Probability, exponential distribution Eye movements, 38, 318 Falmagne, J-C., 86 Fast Fourier transform, 122, 124, 126, 129, 150, 153, 157, 166, 184 Feedback, 65±66, 257±260, 262, 264±266, 269 positive, 45 Feller, W., 100, 102, 106, 119, 168, 247, 253 Feurstein, M. C., 156±157 Feynman, R. P., 4, 25±26, 29, 53, 55, 57±58, 61, 99, 111, 201, 314 Finite state machine, 33, 35, 38, 40, 42, 72± 73, 79 Fitness. See Evolution, ®tness FitzHugh-Nagumo model. See Model neuron, FitzHugh-Nagumo FitzHugh, R. A., 273, 280±281, 306 Fluctuations, 5±7, 24, 29, 32, 47, 87, 90±91, 107, 110, 115, 117, 119, 122, 124±126, 129±132, 135±136, 139±140, 146±148, 150, 158±162, 167, 169±170, 173, 175± 177, 182, 296±297 Fourier analysis, 305 Fourier transform, 122, 126, 150, 157, 165± 166, 184 Fractal, 110, 213±214, 242 Fraser, J. T., 19, 29 Freeman, W. J., 237±240, 242±243 Frequency, 15, 29, 32, 122, 124, 126, 129± 130, 132, 135±136, 147, 155, 159, 166± 167, 171, 177±178, 183±188, 196, 238, 244, 260, 278, 297, 300, 303, 317, 320

Index Frey, B., 138±139, 141, 144, 213, 235 Fries, P., 318, 320 Function, iteration of, 138, 156±157, 202± 205, 208±210, 216, 269, 305 Functional magnetic resonance imaging, 299 Gambling, 99±102, 106, 245 Gaussian distribution. See Probability, Gaussian distribution Gene. See Evolution, genes General relativity, 55, 61 Gilden, D. L., 136±139, 141, 143±144, 235 Gluckman, B. J., 199 Goss-Custard, J. D., 252 Gottman, J. M., 9, 15±16, 235 Graph theory, 4, 22, 66, 72±75, 78, 86, 121, 147, 151, 173, 182, 185, 196, 199, 203, 210, 219, 224, 240, 255, 265, 279, 288, 307, 309 directed graph, 72, 74±75 weighted graph, 73, 75, 79 Grassberger, P., 230, 240 Greenwood, P. E., 74, 99, 141, 170, 173, 179, 182, 187 Gregson, R. A. M., 7, 69, 89, 229, 308 Haken, H., 21, 282 Hall, E. T., 17±19, 29 Herrnstein, R. J., 247±248 Heuristic, 266, 269, 271±272 HKB model. See Model, HKB Hodgkin, A. L., 194, 273, 280±281 Hodgkin-Huxley model. See Model neuron, Hodgkin-Huxley Hurst exponent, 131 Huxley, A. F., 194, 273, 280±281 Implicit memory. See Memory, implicit In®nite regress, 9, 11 Information, 6±7, 10, 13, 29, 33, 37±38, 46± 47, 55, 58±59, 62, 72, 74±75, 85, 107, 116, 141, 145±147, 150, 178, 182±183, 187, 193, 196, 214, 238, 239, 257±58, 265, 283, 287±288, 291, 295, 297±298, 300, 313± 314, 316, 321, 323 Instability, 20±21, 23, 51, 158, 178, 251, 253 Integrate-and-®re neuron. See Model neuron, integrate-and-®re Integration, 11, 24, 64±65, 90, 108, 119, 179, 197±198, 306, 315, 317, 320 Interneuron, 183, 193, 196±97 James, W., 28±29, 282, 315 Jaynes, J., 315, 317, 323 Jourdain, R., 19±20

351 Kahneman, D., 29, 40, 250, 257, 323 Kareev, Y., 257, 260 Kau¨man, S. A., 33, 55, 155, 213, 291 Kelso, J. A. S., 7, 20±22, 50, 282 Kubovy, M., 249±251 Langevin equation, 117±118, 216 Language, 15, 46, 55, 58, 99, 285, 291, 323 Lashley, K. S., 11±12, 28, 237, 240, 311 Law general system, 49 of large numbers, 246±247 matching, 247±248 power, 4, 87, 97, 157, 159 statical, 4 stationary, 79, 103±106 Learning, 9, 12, 14, 21, 38, 41±42, 58, 84, 91, 238±239, 266, 269, 300, 309, 311± 312 Linear regression, 97±98, 115, 229±230 Lockhead, G. R., 250, 255, 261±262, 264 Logic, 40, 46, 234, 312±313 Logistic. See Di¨erence equation, logistic Logothetis, N. K., 317±318 Longtin, A., 196, 199, 281 Luce, R. D., 85 Lumer, E. D., 321 McDonald, J. J., 151, 153 Magic, 1±2, 4 Magnetism, 57±58 Magnetoencephalogram, 147±148, 298±299, 318, 320, 324 Mallon, M. W., 136±138 Mandelbrot, B., 124, 165, 171, 173, 213 Markov chain, 13, 33, 35, 58, 106, 250±251, 257 Markov process, 13, 106, 108 Markov property, 106 Master equation, 108, 110, 118 Mathematics, 7, 13, 15, 57±58, 73, 87, 99, 103, 113 Maxwell, J. C., 56±58 Meme, 53, 315, 323 Memory, 12±13, 19±20, 26, 28±29, 38, 40± 43, 46, 55, 63±64, 67±68, 131, 138±139, 141, 143, 146, 160, 162±163, 235±236, 252±253, 260±261, 269, 271±272, 278± 279, 285, 287, 303, 313 explicit, 63±64, 261, 291 implicit, 63±64 sequence, 42 Mesoscopic description, 107±108, 110 Meter, 19±20 Method of delay coordinates, 210 Metzger, M. A., 266, 269, 271 Microgenesis, 38, 41

352 Model deterministic, 82, 84, 87, 235±236 dynamical, 8, 43, 64, 66±69, 71±72, 74±75, 78±79, 82, 86±87, 89, 273, 283, 285, 292± 294, 306, 309, 314 empirical, 82, 84, 97±98 HKB, 21±22, 24, 282 neural network, 79, 309, 311±312 statical, 62±63, 67±69, 71, 81±82, 85±86 stochastic, 84, 86±87, 99±100, 103, 233 structural, 71, 73, 78 theoretical, 82, 84±85, 97 Model neuron, 160±162, 186, 195±199, 208 FitzHugh-Nagumo, 281 Hodgkin-Huxley, 273, 280±281 Integrate-and-®re, 160, 194±198, 321 Molnar, M., 243±244 Morrison, F., 65, 81±83, 87, 202 Moss, F., 175, 178, 186±187, 190±191 Moving average, 90±92, 94, 96. See also ARIMA Musha, T., 132±133, 136, 146±147 Music, 2, 4, 12, 19±20, 24, 63, 75, 135±136, 238, 287, 300 Nagumo, J., 194, 281 Natural selection. See Evolution, natural selection Neural circuit, 239 Neural network, 79, 160, 163, 208, 301, 308±309, 311±314, 321 Neural network model. See Model, neural network Neuringer, A., 112, 257±259, 263±266, 269 Neural correlate of consciousness. See Consciousness, neural correlate of Neuron, 11±13, 42, 46, 79, 145±148, 150± 151, 153, 160, 162±163, 169, 175, 178, 183±188, 190, 193±199, 208, 238±240, 242, 278, 280±283, 292, 295±296, 298± 301, 306, 308±309, 311±315, 317±318, 320±321, 323±324 Noise 1/f, 7, 125, 129, 130, 132, 139, 141, 157, 159, 166, 169, 170±171, 173, 233 black, 129±131 brown, 126, 129, 131 characteristic, 115, 118 colored, 116, 119, 126, 129, 131±132, 155 current, 117, 167±168, 170 dynamical, 215±216 electrical, 165, 199, 261 essential, 116 external, 116±117, 160, 215 ignorance, 116 independent, identically distributed, 90, 119, 126, 156, 216, 227, 231±232, 251

Index internal, 116±118, 156, 186, 188, 215, 261 measurement, 215, 226 multiplicative, 110, 157 noncharacteristic, 118±119 pink, 126, 129±132, 139 quantum, 116, 118 shot, 115±116 thermal, 115, 117±118 tuning, 158, 178, 182, 300 water, 184, 186 white, 7, 90±91, 94, 125, 129, 131, 137± 141, 143, 146, 157, 162, 166, 170, 184, 234±235, 242 Normal distribution. See Probability, Gaussian distribution Norwich, K. H., 37, 55, 62 Novikov, E., 145, 148 Numerical integration, 305±306 Nunez, P. L., 295, 300±301, 324 Nyquist frequency, 32 Olami, Z., 160 Olfaction, 237±240, 242 Order, 2±3, 9, 11±12, 27±30, 42, 48, 78, 90± 92, 94, 98, 113, 141, 201, 207, 220, 246, 280, 291, 302, 306±307, 315 serial, 28, 282 temporal, 12, 20, 89, 90 Oscillation, 20, 30, 175, 207, 238, 273±274, 276±278, 280±282, 288, 296, 301±303, 317, 321, 323 global, 301±302 local, 301±302 relaxation, 273, 278±279, 281 sinusoidal, 276 Oscillator, 20, 27, 207±208, 273±274, 276± 278, 280, 300 coupled, 278, 282, 312 damped, 207±208, 274, 276±277, 281, 294, 302 forced, 11, 32, 147, 176±178, 199, 207, 276, 278±281, 300, 312 linear, 274, 277 nonlinear, 208, 277, 296 relaxation, 194, 273, 278, 282±283, 294, 300, 302, 306, 312, 323 unforced, 274, 277±278 Pacut, A., 85 Perception, 1, 8, 20±21, 25, 38, 41±42, 55, 61±62, 64, 84, 135±136, 163, 175, 183, 188, 190±191, 237, 240, 245, 251±252, 261, 287, 297, 309, 323 Periodic orbit, 45, 107, 117, 175±177, 203, 205, 207±208, 291, 307 Periodic point, 104, 203±204, 209 Periodicity, 7, 126, 197, 203±204, 213, 221

Index Phase, 21±24, 111, 199, 204, 207±208, 210, 220±221, 223, 226, 230, 239±240, 242, 265, 269, 271±272, 274, 276±278, 280± 281, 291, 307, 320 anti, 21±23 relative, 22, 155 Phase locking, 278, 281 Phase portrait, 220, 239±240 Phase space, 111, 204, 208, 210, 221, 223, 230, 239, 242, 274, 276, 280, 291, 307 Phrasing, 19±20 Physics, 4, 6, 11, 15, 25±28, 32±33, 50, 53, 55±58, 61±62, 68, 87, 99, 100, 107, 110± 112, 115±116, 130, 145, 147±148, 155, 166±167, 201, 207, 242, 283 Pink noise. See Noise, pink Platt, J., 58 Poisson distribution. See Probability, Poisson distribution Positron-emission tomography, 298±299 Potential barrier, 22, 177 Power law. See Law, power Power spectral density, 122, 129, 166 Power spectrum, 122, 124±126, 129±133, 135±139, 141, 144, 147±148, 150, 153, 155, 157, 159±160, 162, 165±166, 168± 171, 173, 184±185, 188, 190, 210, 213, 219, 233 Precision, 26, 32, 57, 59, 82, 113, 116, 203± 206, 265±266, 269 Predictability, 47, 59, 112, 206, 217, 232, 236, 263±264, 269 Prediction, 4, 7±8, 28, 40, 47, 59, 81, 86, 112±113, 116, 187, 202, 205±207, 210, 217, 221, 223±224, 226, 230, 232, 236, 248, 252, 255±256, 262±264, 266, 269, 272 Priming, 41, 63, 162±163 Probability, 6±7, 13, 15±16, 26, 29, 35±37, 40, 42, 47, 73±74, 84, 86, 99±104, 106± 108, 110, 112, 117±119, 121, 124, 126, 146, 156, 158±159, 165, 168, 179, 187, 192, 206, 210, 213, 227, 231, 236, 244± 253, 255±257, 260, 306, 311, 314 density function, 119, 121, 159, 168, 179, 306 distribution, 6, 13, 16, 47, 74, 84, 100, 102± 103, 106±108, 110, 117±119, 121, 124, 126, 156, 158, 165, 168, 206, 210, 213, 227, 247, 249, 306, 311 distribution function, 118, 121, 165, 179, 181 exponential distribution, 168±170 Gaussian distribution, 84, 103, 121±122, 126, 139, 226±227, 249 matching, 248 Poisson distribution, 121 theory, 99, 101, 119, 179, 256

353 uniform distribution, 119, 121±122, 157, 213 Problem solving, 40 Procaccia, I., 230, 240 Psychology, 4, 7, 13, 54±55, 61±62, 67±68, 73, 84, 86, 89, 115, 137±138, 143, 247, 249, 261±262, 286, 315±316 Psychophysics, 4±7, 55, 62±63, 68±69, 82, 84, 86±87, 96±97, 114, 190, 229, 250, 255, 295, 308 Quantization, 45, 215 Quantum computer. See Computer, quantum Quantum electrodynamics, 26, 49, 111 Quantum mechanics, 6, 53, 55, 61, 116 Quantum wave function, 110±113, 314 Queuing, 166±169 Random walk, 85, 91, 100±102, 106, 126, 170 Randomness, 6±7, 47, 55, 68, 74, 78, 81, 83± 87, 91±92, 94, 96, 99±103, 106, 114, 117, 121±122, 126, 137, 146, 155, 160±161, 165±167, 170±171, 175, 177±179, 182, 184, 186, 188±189, 194, 197, 199, 201, 203±210, 213±216, 226±227, 229±231, 233, 240, 245±247, 250±251, 253, 255± 263, 265, 279, 306, 311, 318 Rapoport, A., 249, 250, 251 Rapp, P. E., 241±243 Rate coding, 300 Rate of change, 26, 51, 56, 64±66, 108, 112, 117, 274 Reentrant loop, 321 Relaxation, 41, 146, 169 Relaxation oscillation. See Oscillation, relaxation Relaxation oscillator. See Oscillator, relaxation Rhythm, 1±4, 17±21, 36, 40, 53, 150, 208, 279, 282, 294 of life, 18 poly, 19 Richard, C. M., 139, 235 Roy, A. K., 9, 15±16, 298 Scheier, C., 233±235, 324 SchroÈdinger equation, 111, 113 Schroeder, M., 129±131, 139, 140 Self-organization, 20, 24, 50, 157±160, 323 Self-consciousness. See Consciousness, self Sensation, 1, 6, 35, 37, 41, 63, 68, 78, 85, 147, 150, 160, 162, 175, 183±187, 190± 191, 197, 238, 240, 242±243, 249±250, 320 Sensitive dependence. See Chaos theory, sensitive dependence

354 Sequence, 9, 11±15, 17, 25, 27±28, 33, 36± 38, 40, 42, 81, 96, 103±104, 106, 150, 205, 221, 245±247, 250±253, 255, 257, 260± 262, 264±266, 269, 271, 294, 303, 313 action, 11±12 random, 96, 245, 247, 250, 255±257, 260± 262 Signal detection theory, 6, 61, 85±86, 191± 192, 248±251, 261 Simonotto, E., 188±190 Singer, W., 146, 198, 283, 287, 300, 317 Skinner, J. E., 243±244 Speech, 2, 4, 28, 62, 135±136, 187, 287, 312± 313 State absorbing, 36 change, 35, 176, 239 cognitive, 14, 324 transition, 13, 38, 42, 108 State vector, 111±112 Stationary law. See Law, stationary Stemmler, M., 160, 192 Stevens, S. S., 4±5, 63, 68, 84, 87, 97 Stochastic model. See Model, stochastic Stochastic process, 7, 79, 89±90, 92, 100, 102±108, 110, 165, 169, 210, 213±215, 221, 226±227, 247, 256, 282, 305 linear, 102, 219, 224, 226, 235 nonlinear, 99, 105, 221, 224, 226, 231, 233 Stochastic resonance, 175, 177±178, 182± 193, 195±197, 199, 281 Strange attractor. See Chaos theory, strange attractor. Structural model. See Model, structural Structure, 12, 17, 19±20, 29, 45, 68, 72, 74± 75, 78±79, 82, 87, 90, 96, 98, 116, 122, 135±136, 144, 161, 226±227, 230±233, 247, 264, 271±272, 276, 286, 288, 299± 300, 303, 309 temporal, 20, 82, 84, 90, 92, 96, 116, 124, 138, 165, 220, 251 Superposition, 112, 170 Surrogate data, 227, 231, 234, 242 Synaptic action density, 295, 301±302 Synchrony, 13, 41, 136, 146, 175, 199, 238±239, 242, 283, 300±301, 317±318, 320±321 System closed, 50 cognitive, 47, 50, 78±79, 112, 114, 129± 130, 273±274, 279, 285 dynamical, 7, 20, 24, 29, 37, 50±51, 55, 59, 108, 110, 112, 121, 126, 129, 132, 156±157, 159, 169, 170, 175, 201±202, 205, 210, 213±217, 219±220, 227, 237, 239±240, 273, 291, 305, 307 open, 50

Index properties, 51 state variables, 50±51, 59, 63, 81, 306 trajectory, 51, 59, 217, 221, 306 System theory, 45, 50 dynamical, 24, 37, 50, 55, 59 general, 45±46, 48, 50, 173 Tang, C., 158±159 Teich, M. C., 156±157 Temporal coding, 283, 312 Theiler, J. T., 227, 241±242 Theory dynamical, 8, 42, 308 evolution, 285, 287, 323 formal, 45, 55±59 optimal foraging, 252 quantum, 32, 110±111 queuing, 167±169 statical, 7 Thinking, 35, 40, 42, 45, 111, 130, 265, 273 Thornton, T., 136±138 Threshold detector, 178±179, 192±193, 198 Thurner, S., 156±157 Time absolute, 26±27 arrow of, 28 biotemporal, 29, 321 discrete, 33, 203 eotemporal, 29, 51, 56, 96, 321 estimation, 137, 139, 229±232, 258 evolution, 71, 75, 108, 111±112, 118, 215, 303, 308 explicit function of, 65 implicit function of, 65±66 measurement of, 30, 51 monochronic, 18±19 nootemporal, 29, 321 period, 22, 27, 29±30, 32, 75, 130, 137, 146, 153, 178, 203±204, 246, 259, 274, 276, 278±279, 288, 292, 309 polychronic, 18±19 reaction, 30, 85±86, 90, 115, 125, 138±139, 143 response, 85 scale, 7, 29, 42, 82, 86, 126, 139, 140, 143, 155, 169±170, 229, 248, 251±252, 285± 286, 298, 312, 315, 320, 324 Time series, 7, 17, 30, 32, 40, 68, 89, 90±94, 96±97, 122, 124±126, 130±132, 135, 137± 141, 147±148, 150, 153, 155, 157, 161, 165±166, 168, 170±171, 184±185, 188, 190, 195, 210, 213±215, 219±221, 223± 224, 226±227, 229±236, 240±242, 244, 258, 279, 288, 296±299, 320 chaotic, 234, 240 nonchaotic, 244 rhythmical, 17

Index Time series analysis linear, 89±91 nonlinear, 219, 230 Timing, 6±7, 30, 38, 40±41, 282, 300, 312 Tononi, G., 283, 315, 320±321 Transition matrix, 13±16, 35, 38, 106, 257 Treisman, M., 251±252, 261 Tschacher, W., 233±235 Turing, A., 45±46, 313 Turing machine, 45±46, 313 Tversky, A., 29, 73, 249±251, 257, 323 Unfolding, 4, 28, 33, 37±38, 41, 50, 67, 74, 78, 84, 86, 100, 102, 108, 117±118, 157, 235, 276, 282, 299, 314 Uniform distribution. See Probability, uniform distribution Usher, M., 160 van der Pol, B., 273, 278 Variability, 6, 81±82, 84±87, 90, 92±94, 97, 104, 115, 117±119, 121, 129, 131, 138±140, 197, 217, 227, 229±231 residual, 68, 98, 139, 229±230 response, 6±7, 86, 116, 229 Variance. See Variability Variety, 17±18, 45±47, 55±56, 58, 84, 119 Vigilance, 292±294 Voss, C., 263±266, 269 Voss, R. F., 112, 135±136 Ward, L. M., 25, 29, 41, 47, 63, 68, 94, 96±97, 114, 137, 139, 141, 151, 170, 173, 179, 182, 190±191, 198, 229±231, 235, 250±251, 253, 255, 261±262, 266, 269, 300 Weaver, W., 13, 46, 99, 100, 245, 247 Wefelmeyer, W., 179 Weighted graph. See Graph theory, weighted graph Weinberg, G. M., 45±46, 59, 62, 69, 75, 87, 114 West, R. L., 266, 269 White noise. See Noise, white Wiesenfeld, K., 158±159, 175

355